Digital Pathology Whole-Slide Imaging for Viral Lesions

Overview and Principles of Digital Pathology Whole-Slide Imaging for Viral Lesions

The integration of digital pathology through whole-slide imaging (WSI) into the diagnostic and investigative framework of veterinary virology represents a paradigm shift in how viral lesions are characterized, quantified, and interpreted. For the veterinary clinical pathologist, the transition from conventional light microscopy to the digital realm is not merely a substitution of medium but an expansion of analytical dimensionality. WSI fundamentally transforms the static, two-dimensional observation of histopathologic change into a dynamic, quantifiable, and computationally tractable data stream. The principles governing its application to viral lesions are rooted in the unique nature of viral pathogenesis: the often subtle, cellular-level cytopathic effects, the multifocal and patchy distribution of lesions, the requirement for precise anatomical localization within complex tissue microenvironments, and the critical need for objective, reproducible assessment to inform diagnosis, prognosis, and therapeutic intervention. This section delineates the core principles and operational framework of employing WSI for the evaluation of viral lesions, establishing the foundational context for the detailed methodologies and applications discussed throughout this article.

The Foundational Workflow: From Tissue to Pixels

The journey of a viral lesion from the patient to a digital whole-slide image for computational analysis is a multi-stage process, each step fraught with potential variables that can influence downstream interpretation. The principles of this workflow demand rigorous standardization to ensure that the digital representation is a faithful and analyzable surrogate for the original glass slide.

The initial phase, tissue processing and staining, is the most critical pre-analytical variable. For viral lesions, the choice of stain is paramount. Hematoxylin and eosin (H&E) remains the cornerstone for general histomorphological assessment, revealing the classic hallmarks of viral infection: inclusion bodies (intranuclear or intracytoplasmic), syncytia formation, cellular swelling (ballooning degeneration), apoptosis, necrosis, and inflammatory cell infiltration. However, the detection of specific viral antigens or nucleic acids is often necessary for definitive diagnosis and understanding of viral tropism. Immunohistochemistry (IHC) and in situ hybridization (ISH) are therefore integral companion techniques to H&E in viral pathology. The quality of the IHC or ISH signal-its intensity, specificity, and signal-to-noise ratio-directly dictates the success of subsequent digital quantification. As demonstrated in studies on Bovine Viral Diarrhea Virus detection in ear-notch specimens, optimized IHC protocols are essential for automated digital image analysis to achieve high sensitivity and specificity [24]. Similarly, the investigation of Wesselsbron virus-induced hepatitis leveraged IHC for T cells (CD3), B cells (PAX5), and histiocytes (Iba1) alongside a double stain for hepatocyte proliferation, illustrating the power of multiplexed immunohistochemical approaches [4]. The principle here is that the digital analysis is only as robust as the chemical signal it is designed to measure.

The scanning process itself introduces its own set of principles. High-resolution scanning at 40x magnification (0.25 µm/pixel) is generally considered the standard for histopathological detail, allowing visualization of nuclear and cytoplasmic features critical for identifying viral cytopathic effects. However, this generates enormous file sizes-a single WSI can exceed 10 GB-creating significant storage and computational bottlenecks. Innovative solutions, such as scanning at lower magnifications (e.g., 5x) and using deep learning super-resolution models to computationally reconstruct 40x images from 5x scans, have been shown to preserve diagnostic accuracy for pathologists and AI systems while reducing storage requirements to less than 2% of the original [11]. This principle of intelligent compression is particularly relevant for large-scale surveillance studies in aquaculture, such as those for Infectious Hematopoietic Necrosis Virus or White Spot Syndrome Virus, where thousands of slides may be scanned.

Furthermore, the issue of z-stack scanning is uniquely relevant for certain viral lesions. Melanocytic lesions, for example, may benefit from multiplanar focusing to accurately identify dermal mitoses [16]. For viral lesions, the presence of intranuclear inclusion bodies or the evaluation of viral plaques in 3D tissue constructs may similarly require depth of field beyond a single focal plane. While standard WSI captures a single plane, multi-layer z-stack acquisition and subsequent focal stacking algorithms can be employed, though this dramatically increases scanning time and data volume. The principle of practical utility must guide this choice; for most routine viral diagnoses, a single high-quality focal plane is sufficient.

Computational Approaches for Lesion Detection and Classification

The core analytical power of digital pathology for viral lesions resides in the application of artificial intelligence (AI) and machine learning (ML). The principles guiding these computational approaches are tailored to the specific challenges of WSI data, including its gigapixel scale, spatial heterogeneity, and the often scarce nature of annotated lesion data.

The primary paradigm for WSI analysis is weakly supervised learning, particularly multiple instance learning (MIL) [6, 7, 12, 15, 18, 20]. This approach addresses the impracticality of obtaining pixel-level annotations from pathologists for every viral inclusion body, focus of necrosis, or infiltrating lymphocyte. In MIL, a whole WSI is treated as a "bag" containing many smaller "instances" (image patches or tiles). The slide-level diagnosis (e.g., "positive for Avian Influenza Virus") is the only label provided during training. The algorithm learns to identify which instances are most predictive of the slide-level label, effectively highlighting the most diagnostically relevant regions without requiring explicit segmentation by a human expert. This principle was elegantly demonstrated in a study predicting necrosis in rat livers, where a weakly supervised model using a Generative Adversarial Network for feature extraction could distinguish spontaneous from treatment-related necrosis, a task mimicking the differentiation of direct viral cytopathology from secondary host responses [6]. For virology, this is invaluable; MIL models can learn to detect subtle, sparse, or multifocal viral lesions that might be missed by a human eye fatigued by scanning a massive tissue section.

A more recent and powerful evolution is contrastive and self-supervised learning (SSL). SSL models are first pre-trained on a large corpus of unlabeled WSIs to learn general, robust visual representations of tissue architecture, such as glandular structures, stromal patterns, and cellular morphology. This pre-training phase does not require any diagnostic labels. Subsequently, the model is fine-tuned on a smaller set of labeled slides for a specific task, like detecting Rabies Lyssavirus Negri bodies or Canine Distemper Virus inclusions [2]. This principle of learning fundamental tissue features before task-specific classification significantly improves performance, especially when dealing with the imbalanced datasets common in virology (e.g., rare viral infections like Nipah Virus in Pigs). SSL models are more robust to staining variability and can generalize better across different laboratories and scanner types [7, 9].

Attention mechanisms within MIL frameworks represent another critical principle. These mechanisms assign a weight (attention score) to each image patch based on its contribution to the final prediction. This not only improves classification accuracy but, crucially, provides a spatial map of diagnostic relevance. As shown in studies on lung cancer and colorectal lesions, attention heatmaps can pinpoint the precise location of tumor cells, which for virology translates to localizing infected cells, foci of viral pneumonia, or the specific glomerular tufts affected in a viral nephritis [5, 10, 12, 18]. This is a profound enhancement over a simple binary positive/negative result; it offers a visual, interpretable explanation for the AI's decision, guiding the pathologist to the most critical areas for verification. The principle of interpretable AI is paramount for clinical acceptance, especially in high-stakes diagnoses of notifiable diseases like African Swine Fever Virus or Classical Swine Fever Virus.

Quantitative Morphometry and the Viral Lesion Phenotype

One of the most transformative principles of digital pathology for virology is the ability to move beyond qualitative description to objective, quantitative measurement of the viral lesion phenotype. This is the domain of pathomics or quantitative histomorphometry.

For decades, pathologists have graded viral lesions using semi-quantitative scales (e.g., mild, moderate, severe lymphocytic infiltration). Digital pathology enables the continuous, precise measurement of thousands of morphometric parameters. This includes:

• Cellular and Nuclear Features: Nuclear area, shape (circularity, eccentricity), chromatin texture, and nucleolar size, which can be used to quantify the degree of karyomegaly or nuclear degeneration induced by viruses like Koi Herpesvirus [5, 21].
• Tissue Architecture: Glandular lumen area, stromal density, and the integrity of tissue boundaries, critical for assessing the destructive nature of viruses causing necrosis or fibrosis.
• Cellular Infiltration Dynamics: The density, spatial distribution, and relative proportions of infiltrating immune cells (T cells, B cells, macrophages) are precisely quantifiable. In the study of Wesselsbron virus hepatitis, automated T cell (CD3+) density was found to be ten times higher than B cell (PAX5+) density and strongly correlated with viral load and serum clinical chemistry markers (AST, bilirubin), providing a quantitative readout of disease severity that far exceeds manual estimation [4]. This principle allows for the creation of a comprehensive "immune landscape" of the viral lesion, offering insights into the balance between protective immunity and immunopathology.
• Viral Load and Distribution: Through IHC or ISH, the area of positive staining, the integrated optical density (a surrogate for antigen amount), and the spatial pattern of viral antigen (focal vs. diffuse) can be objectively measured. Automated algorithms for detecting Bovine Viral Diarrhea Virus antigen in ear notches demonstrated 97.4% sensitivity compared to manual review, showcasing the potential for high-throughput, unbiased screening [24]. The principle is that quantitative metrics are more reproducible and sensitive for detecting subtle changes in disease progression or response to antiviral therapy.

The integration of multi-resolution analysis further refines this principle. A WSI can be analyzed at low magnification to assess overall organ architecture and identify the lesional "center of mass" (e.g., a large area of Fowl Pox Virus proliferative lesions). Then, high-magnification patches can be extracted from those specific regions to analyze cellular details like viral inclusion bodies [13, 19]. This hierarchical approach mimics the pathologist's workflow but executes it with pixel-level precision. Furthermore, the principle of spatial transcriptomics and proteomics is on the horizon for digital viral pathology. By computationally co-registering WSI data from H&E and multiplexed IHC with spatially resolved molecular data (e.g., from GeoMx or MERFISH technologies), we can begin to map viral gene expression or host immune gene activation directly onto the histological morphology, creating a truly multi-omic view of the infection [1, 3].

Addressing the Challenges of AI Interpretability and Bias in Viral Pathology

The deployment of AI for viral lesion analysis is not without significant challenges, and the principles for overcoming them are a critical area of active research. Two primary concerns are interpretability and bias.

Interpretability (or Explainable AI - XAI) is the principle that an AI model's decision must be understandable to a human expert. For a clinician to trust an AI's diagnosis of Rabies Lyssavirus or Canine Parvovirus, they need to know why the algorithm made that call. Attention heatmaps are one solution, but more advanced methods are needed [1, 26]. For example, latent space projection, where the model's internal representation of a WSI is visualized in 2D/3D, can show how the AI clusters different viral lesions, revealing if it is learning biologically meaningful features or confounding artifacts (e.g., tissue folds, air bubbles, stain debris). The integration of pathology knowledge graphs or text prototypes (as in the PAT-MIL framework) guides the learning process by aligning image features with expert-defined pathological concepts (e.g., "intranuclear inclusion," "syncytial cell") [8, 12]. This principle ensures that the AI is not a "black box" but a transparent assistant that amplifies, rather than replaces, the pathologist's expertise.

Bias in AI models is a well-documented problem with severe implications for viral diagnostics. A model trained predominantly on WSI from a single geographic region or a specific host species (e.g., chickens infected with Newcastle Disease Virus in a controlled setting) may fail when applied to WSIs from a different region, a different avian species, or field samples with different tissue processing [1]. The principle of dataset heterogeneity is paramount. Datasets must be multi-institutional and incorporate variability in scanners, staining protocols, host genetics, and viral strains [9, 23]. Out-of-distribution detection, where the AI is trained to flag cases that fall outside its training distribution, is a critical safeguard. Furthermore, the issue of long-tailed or imbalanced datasets is acute in virology. Some viral infections (e.g., Tilapia Lake Virus in a specific region) may have abundant data, while others (e.g., Menangle Virus in pigs) are exceedingly rare. Contrastive learning and synthetic data augmentation (using generative models to create novel but realistic viral lesions) are key principles for mitigating this bias and building robust, generalizable models [2, 14].

The Role of Image Quality, Standardization, and Middleware

The final set of principles concerns the operational infrastructure required to make digital pathology for viral lesions a reality in a diagnostic laboratory. Standardization is the bedrock. The United States Department of Agriculture (USDA), World Organisation for Animal Health (WOAH, formerly OIE), and the Centers for Disease Control and Prevention (CDC) emphasize the need for validated diagnostic tests. The digital WSI pipeline must be validated just as rigorously. This includes standard operating procedures for tissue thickness, staining, and scanning parameters. The introduction of digital slide middleware is an emerging solution to manage quality control (QC) [17]. This software acts as a traffic controller between the scanner and the image management system, allowing for real-time QC of each WSI by a technologist, rejecting those with focus issues, out-of-bounds tissue, or labeling errors before they enter the diagnostic workflow. For viral pathogens, this is crucial; a poorly focused WSI could obscure the critical intranuclear inclusion body of Infectious Laryngotracheitis Virus in a poultry trachea, leading to a false negative [25].

The principle of human-in-the-loop (HITL) remains central. While AI can pre-screen for viral lesions with high sensitivity, human oversight is non-negotiable. A study on cervical biopsies found that an AI system, while useful as a "safety layer," could both undercall and overcall glandular lesions, and only expert review could correct these high-consequence errors [22]. For viral diseases, where a misdiagnosis could lead to the culling of an entire flock or herd (e.g., Highly Pathogenic Avian Influenza), a confirmatory review by a boarded veterinary pathologist is mandatory. The AI serves to speed up the review, reduce fatigue-related misses, and provide quantitative data, but the final diagnostic responsibility rests with the human expert. The principle of augmented intelligence, where AI assists the pathologist, is far more pragmatic and effective than the goal of complete automation for complex, context-dependent viral lesions.

Protocol and Methodology for WSI Acquisition and Analysis of Viral Tissue Samples

The translation of viral histopathology from conventional glass slide microscopy to digital whole-slide imaging (WSI) requires a meticulously standardized protocol that addresses the unique biological and technical challenges presented by viral infections. Unlike neoplastic or degenerative conditions, viral lesions often manifest as subtle, patchy, or multifocal alterations within a background of preserved tissue architecture, necessitating acquisition parameters and analytical workflows specifically optimized for pathogen detection. This section delineates the comprehensive methodology for WSI acquisition and computational analysis of viral tissue samples, integrating principles from digital pathology validation studies [16, 17, 31] with emerging artificial intelligence (AI) frameworks [1, 5, 7] that have demonstrated particular utility in infectious disease diagnostics.

Pre-Analytical Considerations and Tissue Preparation

The foundation of any reliable WSI-based viral diagnosis rests upon rigorous pre-analytical standardization. Formalin-fixed, paraffin-embedded (FFPE) tissue remains the gold standard for archival viral pathology, as it preserves both morphological detail and antigenic integrity for downstream immunohistochemical (IHC) confirmation [4, 24]. However, for RNA viruses-including Infectious Hematopoietic Necrosis Virus, Viral Hemorrhagic Septicemia Virus, and Rabies Lyssavirus-fixation duration must be carefully controlled, as prolonged formalin exposure can crosslink nucleic acids and compromise downstream molecular assays. The recommended fixation time for FFPE blocks intended for combined histological and molecular analysis is 24-48 hours at neutral pH, with tissue thickness not exceeding 3-4 mm to ensure uniform penetration [4, 21].

Section thickness for viral WSI must balance optical resolution with the need to visualize intracellular inclusion bodies, syncytia, and other cytopathic effects (CPE). Standard 4-5 μm sections are adequate for most viral lesions, but thinner sections (2-3 μm) may be required for discriminating viral inclusions from nucleoli or karyorrhectic debris, particularly in infections caused by Canine Distemper Virus, Feline Herpesvirus 1, or Marek's Disease Virus [16, 32]. The use of adhesive-coated slides is strongly recommended to prevent tissue detachment during the multiple antigen retrieval steps required for multiplexed IHC protocols targeting viral antigens and immune cell markers simultaneously [3, 4].

Scanning Parameters and Quality Assurance

The selection of scanning magnification and resolution constitutes a critical determinant of diagnostic accuracy for viral tissue samples. While 20× objective scanning (0.5 μm/pixel) is generally sufficient for screening parenchymal lesions, 40× scanning (0.25 μm/pixel) is strongly recommended for the identification of intranuclear or intracytoplasmic viral inclusions, which may measure only 1-5 μm in diameter [16, 29]. Studies comparing WSI to glass slide diagnosis for melanocytic lesions-which share with viral infections the requirement for high-resolution cytological detail-demonstrated that 40× scanning achieves diagnostic concordance rates exceeding 95% [16, 31]. For aquatic viral pathogens such as White Spot Syndrome Virus and Decapod Iridescent Virus 1, where inclusion bodies may be subtle or pleomorphic, extended depth-of-field imaging through z-stack acquisition may be necessary to capture the full three-dimensional morphology of infected cells [16]. However, routine implementation of z-stack scanning for all viral cases is not recommended due to the significant increase in file size (typically 5-10× larger than single-plane scans) and the marginal improvement in diagnostic accuracy observed for most lesions [16].

Quality assurance protocols must be embedded within the scanning workflow to detect pre-analytical and analytical defects that disproportionately affect viral tissue samples. A multicenter European study of melanocytic lesions-a model system for challenging WSI interpretation-found that 46.3% of scanned slides exhibited technical defects, with 70.1% classified as analytical (focus, stitching, or color artifacts) and 29.9% as pre-analytical (tissue folding, tearing, or incomplete sectioning) [23]. For viral tissues, which are often submitted from field settings where fixation and processing may be suboptimal, the defect rate may be even higher. Implementation of digital slide middleware-a software layer positioned between the scanner and image management system-enables real-time quality control, allowing pathology technicians to accept or reject scans within seconds of acquisition and to trigger automated rescans for failed slides [17]. This middleware can also integrate rules-based notifications for stains requiring polarization (e.g., Congo red for amyloid in chronic viral infections) and flag slides with insufficient tissue coverage or labeling errors [17].

Histochemical and Immunohistochemical Staining Protocols for Viral WSI

Hematoxylin and eosin (H&E) staining remains the primary screening modality for viral lesions, providing the architectural context within which CPE are identified. However, the diagnostic sensitivity of H&E alone for viral infections is limited, particularly for agents that induce minimal or nonspecific histopathological changes. For example, Bovine Viral Diarrhea Virus infection in persistently infected calves may present with no discernible microscopic lesions, yet immunohistochemical detection of viral antigen in ear-notch specimens achieves 97.4% sensitivity and 89.4% specificity when coupled with WSI-based digital analysis [24]. This integration of IHC with WSI and automated image analysis software-specifically the HALO platform (Indica Labs)-has transformed BVDV surveillance by reducing technician review time from hours to minutes while eliminating the visual fatigue that contributes to false-negative readings [24].

For comprehensive viral lesion characterization, multiplexed IHC protocols are increasingly employed to simultaneously visualize viral antigens, cellular markers of infection (e.g., CD3 for T cells, PAX5 for B cells, Iba1 for histiocytes), and proliferation indices (Ki67) within the same tissue section [3, 4]. The Wesselsbron virus hepatitis model demonstrated that digital quantification of CD3+ T cell density, Iba1+ histiocyte infiltration, and hepatocyte proliferation index (arginase 1/Ki67 double staining) correlated significantly with viral load measured by RT-qPCR and with serum hepatic injury markers including aspartate transferase and bilirubin [4]. This approach revealed that histiocyte density exhibited the strongest correlation with viral load, underscoring the dominant role of macrophages in Wesselsbron virus pathogenesis-a finding that would have been difficult to appreciate through conventional semiquantitative scoring [4].

Computational Analysis Pipelines for Viral Lesion Detection

The application of AI and machine learning to viral tissue WSI has advanced from proof-of-concept to validated clinical tools, driven by the development of weakly supervised learning frameworks that circumvent the need for labor-intensive pixel-level annotations [6, 7, 18]. In veterinary viral pathology, where annotated datasets are scarce and lesion heterogeneity is high, weakly supervised approaches are particularly advantageous. These methods train on slide-level labels (e.g., "infected" vs. "uninfected") and automatically identify discriminative regions of interest through attention mechanisms or clustering-constrained multiple instance learning [6, 7, 15].

The PathologAI system, developed for rat liver toxicity assessment and applicable to viral hepatitis models, employs a generative adversarial network to preprocess WSI tiles and ensemble convolutional neural network classifiers to predict necrosis with 87% accuracy on control slides and 83% on slides with spontaneous lesions [6]. Critically, this system demonstrated the ability to discriminate treatment-related from spontaneous lesions-a distinction that is directly analogous to differentiating viral-induced from background pathology in field samples [6]. For viral hepatitis caused by Classical Swine Fever Virus or African Swine Fever Virus, such discrimination is essential for accurate diagnostic classification.

More advanced frameworks integrate self-supervised vision transformers for feature extraction with novel guided attention mechanisms that handle the heavily imbalanced datasets typical of viral pathology, where lesional tissue may constitute less than 1% of the total WSI area [7]. These attention-guided models achieved a 38% improvement in AUC over conventional methods for rare lesion detection in rat livers [7], suggesting that similar architectures could be effective for detecting early or focal viral lesions in conditions such as Porcine Reproductive and Respiratory Syndrome Virus interstitial pneumonia or Infectious Bursal Disease Virus lymphoid depletion.

Spatial Quantification and Multimodal Integration

Viral infections are inherently spatial processes, with the distribution of infected cells within the tissue microenvironment determining both pathogenesis and host response. Digital pathology enables precise spatial quantification of viral antigen distribution, immune cell infiltration, and tissue damage at scales ranging from individual cells to whole organs [3, 28]. The pancreatic histopathology study of type 1 diabetes progression-a model system for understanding viral-triggered autoimmunity-employed semi-automated image analysis of approximately 25,000 islets to reveal that islet size and spatial homogeneity are organized in an islet size-contingent manner, and that histopathological correlates of disease progression are detectable even at preclinical stages [3].

For aquatic viral infections affecting aquaculture species, the spatial distribution of lesions within target organs provides critical diagnostic and prognostic information. Infectious Salmon Anemia Virus infection in Atlantic salmon produces characteristic hemorrhagic and necrotic lesions in the liver and kidney that can be quantified using WSI-based morphometry to establish lesion severity scores that correlate with clinical outcome. Similarly, Tilapia Lake Virus induces syncytial hepatitis with pathognomonic intracytoplasmic inclusion bodies, the distribution and density of which can be mapped across entire tissue sections using deep learning-based segmentation to differentiate infected from uninfected hepatocytes [21].

Multimodal integration-combining WSI-derived morphological features with molecular, serological, and clinical data-represents the frontier of viral lesion analysis [1, 27, 30]. For chronic hepatitis B-associated liver fibrosis, a deep learning model integrating H&E and Masson-stained WSI with clinical features achieved an AUC of 0.741 for predicting fibrosis regression after antiviral therapy, with gradient-weighted class activation mapping revealing that the model focused on hepatocyte degeneration, disorganized hepatic cords, and thick-bridging fibrous septa as key predictors [27]. This multimodal framework can be adapted for viral hepatitis caused by Duck Hepatitis A Virus, Goose Parvovirus, or Rabbit Hemorrhagic Disease Virus 2, where the combination of histopathological features with viral load and serological status may improve prognostic accuracy.

Validation and Regulatory Considerations

The translation of WSI-based viral lesion analysis from research to routine diagnostic or regulatory use requires rigorous validation against established reference standards. For regulatory toxicologic pathology, the National Toxicology Program has demonstrated that WSI-based peer review achieves concordance with glass slide review ranging from 74% to 100% (median 86%), with intra- and interobserver variation not influencing study conclusions [33]. However, subtle viral lesions-particularly those involving small inclusions within large tissue sections or lesions at the limits of optical resolution-may show lower concordance and warrant continued use of glass slides for primary diagnosis [25, 33].

For infectious disease surveillance programs, validation protocols must explicitly address the sensitivity and specificity of WSI-based detection relative to gold-standard methods (virus isolation, RT-qPCR, or IHC on glass slides). The BVDV ear-notch IHC study achieved 97.4% sensitivity and 89.4% specificity using WSI with automated digital analysis, but noted that false negatives occurred predominantly in slides with weak staining or high background, which can be mitigated through optimization of antigen retrieval protocols and staining quality control [24]. Similarly, the use of AI for cervical biopsy diagnosis at the CIN2+ threshold-which is directly relevant to papillomavirus-associated lesions in both human and veterinary medicine-demonstrated that AI systems can detect most clinically important lesions but may undercall or overcall glandular lesions, necessitating expert pathologist review for high-consequence misclassifications [22].

The implementation of digital pathology for viral tissue analysis in regulatory settings (e.g., WOAH-listed diseases such as Foot-and-Mouth Disease Virus, African Swine Fever Virus, or Highly Pathogenic Avian Influenza Virus) requires compliance with strict quality management systems. The middleware-based quality control approach described previously [17] can be extended to include automated detection of scanner drift, color calibration discrepancies, and tissue coverage adequacy, with all metadata logged for audit trails. Furthermore, the European experience with establishing a second-opinion network for rare melanocytic lesions highlights the importance of standardized scanning protocols across institutions to minimize inter-center variability in file size, resolution

Molecular Pathogenesis and Histomorphologic Correlates of Viral Infections

The interrogation of viral infections through digital pathology whole-slide imaging (WSI) represents a transformative convergence of molecular virology, quantitative histomorphometry, and computational pathology. The capacity to systematically correlate viral replication strategies with tissue-level histomorphologic alterations at unprecedented scale and resolution has fundamentally redefined the analytical frameworks available to the veterinary clinical pathologist. This section synthesizes the current understanding of molecular pathogenesis across diverse viral families, emphasizing the spatially explicit, quantitative histomorphologic correlates that are now accessible through WSI-based analytical pipelines.

From Descriptive Histology to Quantitative Pathogenesis

Traditional viral pathology has relied upon qualitative, observer-dependent descriptions of cytopathic effect, inflammatory topography, and tissue tropism. The advent of high-resolution WSI, coupled with machine learning-driven image analysis, has enabled a paradigm shift toward the objective, reproducible quantification of lesion-specific parameters. Studies employing weakly supervised deep learning frameworks on WSI from animal models have demonstrated the capacity to discriminate treatment-related necrosis from spontaneous background lesions, achieving classification accuracies exceeding 80% in rat liver toxicity studies [6]. This capability is directly translatable to viral pathogenesis research, where the discrimination between virus-induced cytopathic effect and secondary inflammatory tissue damage is often diagnostically decisive.

The molecular pathogenesis of viral infection is fundamentally a narrative of host-pathogen interactions unfolding across spatial and temporal dimensions. Digital pathology, when integrated with immunohistochemical and in situ hybridization modalities, permits the precise mapping of viral antigen distribution relative to histomorphologic injury. A landmark study employing machine learning-driven digital histopathology for Wesselsbron virus (WSLV)-induced hepatitis in ewes and lambs demonstrated that digitally quantified parameters-including lymphohistiocytic infiltration density and hepatocyte proliferation index-positively correlated with viral load as determined by RT-qPCR and serum hepatic injury markers including aspartate transferase and bilirubin [4]. This establishes a direct, quantifiable bridge between viral replication kinetics and histomorphologic outcome, a linkage that is mechanistically essential for understanding virulence determinants.

Viral Tropism and Cytopathic Mechanism: Histomorphologic Signatures

The histomorphologic correlates of viral infection are determined by the specific molecular mechanisms of viral entry, replication, assembly, and egress, as well as the host cellular response. Viruses exhibiting epitheliotropism, such as Feline Herpesvirus 1 and Canine Adenovirus 2, typically produce intranuclear inclusion bodies, syncytial cell formation, and epithelial necrosis with associated neutrophilic or lymphocytic inflammation. The WSI-based quantification of inclusion body density, syncytial cell frequency, and epithelial denudation area provides objective metrics for assessing virulence and therapeutic response. Conversely, viruses targeting hematopoietic or lymphoid tissues-including Feline Leukemia Virus, Avian Leukosis Virus, and Bovine Leukemia Virus-induce hyperplasia, neoplasia, or depletion of target cell populations, alterations that are amenable to quantitative assessment through nuclear segmentation and cell classification algorithms applied to WSI.

For neurotropic viruses, the histopathologic landscape is particularly complex, encompassing neuronal necrosis, gliosis, perivascular cuffing, and demyelination. Canine Distemper Virus exemplifies a pathogen whose molecular pathogenesis involves hematogenous dissemination followed by neural invasion, leading to demyelinating leukoencephalitis characterized by intranuclear and intracytoplasmic inclusion bodies within neurons and glial cells. The WSI-based quantification of demyelinated plaque area, astrocyte hypertrophy, and microglial nodule density offers a powerful approach for correlating viral antigen distribution with the severity and topography of neuropathology. Similarly, West Nile Virus in Birds and West Nile Virus in Horses produce a predominantly lymphoplasmacytic encephalomyelitis with neuronal degeneration and neuromophagia, lesions that can be systematically mapped across entire brain sections using digital pathology platforms. The integration of multiplexed immunohistochemistry with WSI-as demonstrated in studies of type 1 diabetes progression, where ~25,000 islets were analyzed for immune cell composition and spatial relationships [3]-provides a methodological template for deep immunophenotyping of virus-induced inflammatory lesions.

Hepatotropic Viruses: A Case Study in Digital Histomorphometry

The liver represents a frequent target of viral infection across multiple species, and the histomorphologic spectrum of viral hepatitis-ranging from acute hepatocellular necrosis to chronic active hepatitis, fibrosis, and cirrhosis-is ideally suited for quantitative WSI analysis. The aforementioned WSLV study [4] provides a paradigmatic example: digital quantification of CD3+ T cell density, PAX5+ B cell density, Iba1+ histiocyte density, and hepatocyte proliferation index (arginase 1/Ki67 dual staining) revealed that the density of Iba1+ histiocytes exhibited the strongest positive correlation with hepatic viral load. This observation underscores the prominent role of macrophages and histiocytes in the host response to flaviviral hepatitis and demonstrates the power of automated, whole-slide cellular phenotyping to uncover pathogenesis-relevant cell populations that might be overlooked in conventional histologic assessment.

Other hepatotropic viruses of veterinary importance, including Bovine Viral Diarrhea Virus, Classical Swine Fever Virus, Duck Hepatitis A Virus, and Avian Hepatitis E Virus, induce characteristic patterns of hepatocellular injury, apoptosis, and inflammation. For BVDV, the detection of persistently infected calves relies upon immunohistochemical detection of viral antigen in ear-notch specimens. A recent study employing HALO digital image analysis for automated BVDV antigen detection achieved 97.4% sensitivity and 89.4% specificity compared to manual assessment, with significantly reduced personnel time for slide review [24]. This represents a practical application of digital pathology for high-throughput viral surveillance, illustrating how molecular pathogenesis (persistent viral antigen expression in epithelial and lymphoid tissues) can be exploited for diagnostic screening through automated WSI analysis.

Vascular and Endothelial Tropism: Hemorrhagic and Edematous Lesions

Several viral pathogens exhibit prominent tropism for vascular endothelium, leading to increased vascular permeability, hemorrhage, and edema. African Swine Fever Virus targets macrophages and endothelial cells, producing widespread hemorrhagic necrosis of lymphoid tissues, with intense perivascular erythrocyte extravasation and fibrin deposition. The WSI-based quantification of hemorrhagic area, fibrin deposition density, and endothelial cell infection index provides objective correlates of disease severity. Similarly, Equine Arteritis Virus produces a panvasculitis with medial necrosis and perivascular edema, lesions that are amenable to automated morphometric analysis of vessel wall thickness and perivascular fluid accumulation. For aquatic species, Viral Hemorrhagic Septicemia Virus and Infectious Salmon Anemia Virus induce severe vascular injury with widespread petechiation and organ necrosis, histomorphologic features that can be systematically quantified through digital pathology pipelines validated for fish tissue.

The molecular mechanisms underlying endothelial tropism involve viral glycoprotein interactions with host cell receptors, activation of endothelial signaling pathways, and disruption of tight junction integrity. The histopathologic correlate-perivascular hemorrhage, edema, and leukocyte margination-can be precisely measured using segmentation algorithms applied to WSI. Deep learning models trained to distinguish between hemorrhage, edema, and necrosis in tissue sections have shown high accuracy in preclinical toxicology studies [6, 7], and these same architectures are directly applicable to the characterization of viral hemorrhagic syndromes.

Respiratory Viral Infections: Remodeling the Airway Microenvironment

Respiratory viruses, including Avian Influenza Virus, Swine Influenza A Virus, Equine Influenza A Virus, Infectious Bronchitis Virus, and Bovine Respiratory Syncytial Virus, induce a spectrum of histomorphologic alterations spanning from epithelial degeneration and necrosis to mucosal hyperplasia, squamous metaplasia, and fibroplasia. The molecular pathogenesis involves viral replication within respiratory epithelial cells, triggering innate immune responses characterized by neutrophil and macrophage infiltration, mucus hypersecretion, and impairment of mucociliary clearance. WSI-based analysis of epithelial thickness, goblet cell density, luminal neutrophil accumulation, and submucosal gland hyperplasia provides quantitative metrics for comparing virulence across viral strains and evaluating therapeutic interventions.

For avian species, the respiratory tract presents unique histomorphologic features, including the presence of air capillaries and parabronchi rather than alveoli. Avian Metapneumovirus and Turkey Coronavirus produce characteristic lesions of the upper respiratory tract, including rhinitis, sinusitis, and tracheitis with epithelial deciliation, mononuclear cell infiltration, and lymphoid hyperplasia. Digital pathology platforms capable of handling the large tissue sections required to capture the entire respiratory tract of poultry-from nasal turbinates to lung-are essential for comprehensive pathogenesis studies. The integration of WSI with quantitative image analysis permits the objective assessment of lesion severity and distribution, facilitating comparisons across experimental groups in vaccine efficacy and challenge studies.

Enteric Viral Infections: Villous Atrophy, Crypt Hyperplasia, and Inflammatory Topography

Enteric viruses, comprising rotaviruses, coronaviruses, astroviruses, caliciviruses, and parvoviruses, induce a characteristic histomorphologic pattern of villous blunting and fusion, crypt hyperplasia, and lymphoplasmacytic infiltration of the lamina propria. The molecular pathogenesis typically involves viral replication within mature enterocytes at the villous tip, leading to cell death, villous collapse, and compensatory crypt cell proliferation. Canine Parvovirus, Feline Panleukopenia Virus, and Porcine Epidemic Diarrhea Virus exemplify viruses producing severe enteropathy with extensive villous necrosis and crypt destruction.

The WSI-based morphometric analysis of villous height-to-crypt depth ratio, as a continuous quantitative variable, provides a sensitive and reproducible measure of enteric injury. Automated segmentation of intestinal villi and crypts using convolutional neural networks-similar to those validated for testicular histology [34] and kidney tissue compartments [35]-enables high-throughput assessment of enteric lesion severity across entire intestinal sections. This approach is particularly valuable for assessing the efficacy of antiviral interventions and vaccines in experimental challenge models, where subtle differences in villous architecture may be missed by subjective histologic grading. Furthermore, multiplexed immunohistochemical staining combined with WSI allows simultaneous visualization of viral antigen, cellular markers of proliferation (Ki67), apoptosis (cleaved caspase-3), and immune cell subsets, providing a spatially explicit map of the pathogenic process.

Lymphoid Organ Pathology: Depletion, Hyperplasia, and Follicular Architecture

Viral infections frequently target lymphoid organs, producing histomorphologic alterations that reflect the underlying molecular interaction between virus and immune cells. Infectious Bursal Disease Virus causes profound lymphoid depletion within the bursa of Fabricius, with follicular necrosis, cystic transformation, and interfollicular fibrosis. Porcine Circovirus 2 induces lymphoid depletion and histiocytic infiltration in lymph nodes and spleen, a hallmark of postweaning multisystemic wasting syndrome. Marek's Disease Virus produces lymphoproliferative lesions with T cell lymphoma formation in visceral organs, nerves, and skin.

The WSI-based analysis of lymphoid follicle size, density, and cellular composition provides quantitative correlates of immune status and viral pathogenesis. Automated segmentation of follicular versus interfollicular compartments, coupled with nuclear classification algorithms, enables the calculation of B cell/T cell ratios, germinal center area, and mantle zone thickness. These metrics are directly relevant to understanding the immunopathogenesis of viruses that modulate lymphoid architecture, including immunosuppressive retroviruses such as Feline Immunodeficiency Virus and Equine Infectious Anemia Virus. The application of deep learning models trained to detect and classify lymphoid follicles in WSI, similar to approaches validated for renal pathology and cancer diagnosis [35, 36], offers a pathway toward standardized, objective assessment of lymphoid pathology in viral disease.

The Inflammatory Microenvironment: Cellular Composition and Spatial Architecture

The host inflammatory response to viral infection is a critical determinant of both viral clearance and immunopathology. Digital pathology, when combined with multiplexed immunohistochemistry and computational analysis, permits the detailed characterization of the inflammatory microenvironment at single-cell resolution across entire tissue sections. The study of WSLV-induced hepatitis [4] demonstrated that T cell density was approximately 10-fold higher than B cell density in infected animals, with strain-dependent differences in the magnitude of infiltration. This level of quantitative immunophenotyping, achieved through automated cell counting on WSI, provides insights into the cellular mechanisms of viral clearance and tissue injury that are not accessible through conventional histology.

For viruses that establish persistent or latent infections, including Bovine Herpesvirus 1, Equine Herpesvirus 1,

Clinical Application and Diagnostic Performance in Viral Disease Detection

The translation of digital pathology whole-slide imaging (WSI) from research platforms into frontline veterinary diagnostic practice represents a paradigm shift of profound consequence for the detection and characterization of viral diseases. Unlike neoplastic pathology, where computational tools have been largely validated for grading and classification tasks, the viral disease diagnostic landscape presents unique challenges: the often-subtle nature of virus-induced cytopathic effect, the need for spatial contextualization of antigen within tissue microenvironments, the inherent low prevalence of positive samples in surveillance contexts, and the critical requirement for rapid turnaround during outbreak responses. The clinical application of WSI-based digital pathology in viral disease detection thus demands not merely technical equivalency to conventional light microscopy but demonstrable added value in sensitivity, specificity, reproducibility, and workflow efficiency. This section critically examines the current evidence base, drawing upon a spectrum of applications from aquatic, avian, livestock, companion animal, and wildlife virology paradigms.

Quantitative Immunohistochemical Detection of Viral Antigens: The Paradigm of Bovine Viral Diarrhea Virus Surveillance

The most compelling evidence for the clinical utility of digital pathology in viral disease detection comes from surveillance programs for persistently infected (PI) animals, where the diagnostic challenge is not the identification of florid lesions but the consistent detection of low-level antigen expression across large tissue volumes. The application of WSI coupled with automated image analysis for Bovine Viral Diarrhea Virus antigen detection in ear-notch immunohistochemistry (IHC) specimens serves as an exemplary model. In a rigorous evaluation spanning 518 slides containing 2,884 individual ear notches, a digital pathology workflow utilizing HALO image analysis software achieved a sensitivity of 97.4% and specificity of 89.4% against the gold standard of direct human assessment [24]. Critically, the time required for personnel to operate the software and organize results was significantly shorter than manual microscopic examination, demonstrating that digital pathology can address the fundamental bottleneck in BVDV surveillance-the disproportionate effort expended on negative samples that constitute the vast majority of submissions [24]. This is a recurring theme in viral diagnostics: the low-prevalence surveillance context makes traditional manual screening inefficient, whereas algorithmic prescreening can dramatically reduce pathologist fatigue and improve diagnostic focus. The authors of this landmark study noted that the benefits of reduced potential for human error and significant time savings for both technicians and pathologists were evident even as they encountered challenges during the implementation phase, underscoring that the path to clinical integration is iterative rather than instantaneous [24].

Leveraging Machine Learning for Lesion Characterization in Viral Hepatitis: Insights from Wesselsbron Virus

Beyond simple antigen detection, the true diagnostic power of digital pathology in virology lies in its capacity for quantitative, objective lesion profiling. The comprehensive investigation of Wesselsbron virus (WSLV)-induced hepatitis in ewes and lambs provides an illustrative case study of how machine learning-driven digital histopathology can unravel the intricacies of viral pathogenesis while simultaneously serving diagnostic purposes [4]. In this meticulously designed study, Grau-Roma and colleagues employed a pathologist-led digital histopathological workflow on IHC-stained whole-slide images to quantify T cell (CD3⁺), B cell (PAX5⁺), and histiocyte (Iba1⁺) densities, hepatocyte proliferation indices, and viral NS1 antigen distribution. The results were illuminating: infected animals exhibited significantly higher lymphohistiocytic infiltration and hepatocyte proliferation indices compared to mock-infected controls, with T cell densities exceeding B cell densities by a factor of ten [4]. Most importantly for diagnostic performance, the digitally quantified parameters correlated positively with WSLV RT-qPCR results and serum hepatic injury markers (aspartate transferase, bilirubin, adenosine deaminase), demonstrating that digital histopathology reliably detects liver damage and disease severity. The strongest correlation was observed between Iba1⁺ histiocyte density and viral load, a finding that would be exceptionally difficult to derive through conventional semi-quantitative assessment [4]. This study establishes a template for how WSI-based digital pathology can bridge the gap between histomorphological description and quantitative molecular virology, providing a reproducible framework for assessing disease progression in viral hepatitis that could be extrapolated to other hepatotropic viruses such as Rift Valley Fever Virus, Bluetongue Virus, and Avian Hepatitis E Virus.

Telemicrobiology and the Detection of Viral Inclusion Bodies: The Resolution Imperative

The detection of viral inclusion bodies, cytopathic effect, and syncytia formation represents a domain where digital pathology must overcome fundamental optical limitations. The landmark comparison of digital pathology platforms for telemicrobiology applications by Rhoads and colleagues established that the diagnostic accuracy of WSI for identifying viral, bacterial, fungal, and parasitological agents is critically dependent on scanning magnification [29]. Their findings demonstrated that ×83 oil-immersion WSI, ×100 oil-immersion WSI, and digital photomicrographs yielded interpretations that were not significantly different in quality and accuracy from glass slide evaluation. In stark contrast, the standard ×40 WSI scans (the most commonly employed scanning protocol in clinical pathology laboratories) produced interpretations of lower quality and were more likely to be incorrect [29]. This has profound implications for viral diagnostics, where the identification of intranuclear or intracytoplasmic inclusion bodies-hallmarks of infections such as those caused by Canine Distemper Virus (intranuclear and intracytoplasmic eosinophilic inclusions), Feline Herpesvirus 1 (intranuclear Cowdry type A inclusions), and Rabies Lyssavirus (Negri bodies in neurons)-often requires high-magnification, oil-immersion examination. The digital pathology community must therefore recognize that a single scanning resolution does not suffice for all diagnostic contexts; viral cytopathology, in particular, necessitates either z-stack scanning or high-magnification acquisition protocols to maintain diagnostic equivalence with traditional microscopy. This is further supported by the observation that pathologists evaluating melanocytic lesions utilized software focusing options in 7-28% of cases, and that in one instance of nevoid melanoma (a mimic that shares diagnostic challenges with virus-infected cells), z-stack-enabled focusing corrected a misdiagnosis after revealing a dermal mitosis [16]. The principle is clear: multiplanar focusing capacity, whether achieved through z-stack scanning or sophisticated digital refocusing algorithms, is not a luxury but a necessity for viral disease detection.

Addressing Diagnostic Heterogeneity: The Role of Digital Pathology in Multi-Institutional Viral Disease Working Groups

The diagnosis of viral lesions is notoriously subject to interobserver variability, particularly when lesions are subtle, early, or atypical. This heterogeneity becomes critically important in the context of outbreak investigations and international surveillance, where standardized case definitions are paramount. The National Toxicology Program's experience with WSI for pathology peer review offers valuable insights applicable to viral disease diagnostics. Across six pathology working groups and one peer review, the concordance of consensus diagnoses based on WSI versus glass slides ranged from 74% to 100% (median 86%), with the variation not influencing the conclusions of any study [33]. However, this same body of work identified specific scenarios in which WSIs may be suboptimal: evaluation of subtle lesions, large complex lesions, small lesions within large tissue sections, and foci of altered hepatocytes [33]. These limitations are directly transferable to virology contexts. For instance, the detection of early Avian Influenza Virus-induced encephalitis or the identification of sparse African Swine Fever Virus antigen-positive cells in tonsillar tissue of chronically infected pigs may be compromised if digital slide quality is suboptimal. Furthermore, the significant inter-center variability in file sizes and scanning protocols documented in the MELCAYA experience-where 46.3% of WSIs exhibited technical defects, 70.1% of which were analytical in nature-highlights the challenges of obtaining standardized, high-quality WSIs in real-world multisite settings [23]. For viral disease diagnostics, where a single missed diagnosis can have enormous consequences for trade, herd health, or public health, these quality control challenges must be addressed through standardized scanning protocols, robust digital slide middleware for real-time quality assessment, and clear guidelines for when digital slides are acceptable versus when glass slide review is required.

Workflow Integration and the Pipeline for High-Throughput Viral Diagnosis

The practical deployment of digital pathology for viral disease detection requires seamless integration into the diagnostic laboratory workflow. The development of digital slide middleware, as described by Smith [17], provides a critical control point that is difficult or impossible to replicate using native scanner and image management system features alone. Such middleware enables real-time quality accept/reject calls on digital slides from workstations, rescans of rejected slides displayed adjacent to originals, and rules-based notifications ensuring pathologists are aware of digital slide availability for urgent cases-features essential for outbreak response situations where turnaround time is measured in hours, not days [17]. This infrastructure, analogous to the chemistry and hematology middleware that has long been standard in clinical pathology laboratories, is equally essential for virology applications. When combined with AI-driven prescreening algorithms-such as the weakly supervised approaches that have demonstrated 98.8% sensitivity for colorectal lesion detection [37]-the potential for automated triage of viral disease cases becomes tangible. For example, a high-throughput IHC screening pipeline for Porcine Reproductive and Respiratory Syndrome Virus antigen detection in lung tissue could prioritize slides with high probability of positivity, reducing the manual review burden by an order of magnitude while maintaining diagnostic sensitivity.

Persistent Challenges and Path Forward

Despite the clear promise, significant barriers to the widespread clinical adoption of digital pathology for viral disease detection remain. The data scarcity problem is acute: viral lesions are often rare in routine diagnostic submissions, and training robust AI models requires large, diverse, and well-annotated datasets that span multiple viral species, host species, tissue types, and lesion severities. The successful application of self-supervised and contrastive learning strategies to address labeled data insufficiency in other pathology domains [1, 2] offers a potential pathway, but these approaches require careful validation for virology-specific tasks. Furthermore, the biological complexity of viral infections-including mixed infections, concurrent bacterial or parasitic pathogens, and host immune status-dependent lesion variation-presents a level of heterogeneity that may challenge even the most sophisticated machine learning models. The recent demonstration that weakly supervised deep learning approaches can distinguish spontaneous from treatment-related necrosis in rat liver [6] and can classify fibro-osseous lesions of the jaw with accuracy exceeding that of experienced pathologists [20] suggests that similar frameworks could be developed for differentiating viral-induced lesions from those caused by toxins, metabolic disease, or immune-mediated processes. Nevertheless, caution is warranted; the toxicologic pathology community has appropriately noted that digital pathology assessment, while having made great strides, is not yet ready for complete replacement of glass slides in safety assessments, particularly for subtle lesions [25]. This caution applies doubly to viral diagnostics, where the stakes-including international trade restrictions, zoonotic risk assessments, and animal welfare-are extraordinarily high. The path forward demands rigorous, multi-institutional validation studies that compare digital pathology performance against reference standards that include not only conventional microscopy but also molecular diagnostics, virus isolation, and clinical outcome data. Only through such comprehensive validation can digital pathology achieve the diagnostic confidence required to serve as a frontline tool in veterinary viral disease detection.

Integration of Artificial Intelligence and Machine Learning for Viral Lesion Classification

The integration of artificial intelligence (AI) and machine learning (ML) into digital pathology has catalyzed a paradigm shift in the classification of viral lesions, transitioning the discipline from a predominantly descriptive, qualitative endeavor into a quantitative, predictive science [1, 21, 26]. For the veterinary clinical pathologist, the application of AI-driven analysis to whole-slide images (WSIs) of viral infections presents an unprecedented opportunity to enhance diagnostic accuracy, standardize lesion grading, uncover subvisual morphometric signatures, and scale diagnostic capacity to meet the demands of population-level surveillance and outbreak response [4, 30]. This convergence is not merely a technological augmentation but a fundamental re-engineering of how histopathological data are extracted, interpreted, and integrated into clinical and epidemiological decision-making frameworks.

The Core Challenge: From WSI Complexity to Actionable Classification

The fundamental obstacle in classifying viral lesions from WSIs is the immense scale and heterogeneity of the data. A single WSI can contain tens of billions of pixels at 40× magnification, and the pathognomonic features of viral cytopathology-such as inclusion bodies, syncytia, cellular ballooning, or specific patterns of necrosis and inflammation-may occupy only a minute fraction of that total area [6, 18]. This "needle in a haystack" problem is compounded by the tremendous morphological diversity seen across different viral pathogens. For instance, the intranuclear inclusion bodies characteristic of Channel Catfish Virus or Koi Herpesvirus bear little resemblance to the cytoplasmic inclusion bodies of Fowl Pox Virus or Canine Distemper Virus. Furthermore, the same virus can produce vastly different histological patterns depending on host species, tissue tropism, stage of infection, and immune status, as seen with the varied lesions of Avian Influenza Virus across different poultry species or the neurotropic versus pneumotropic forms of Newcastle Disease Virus.

Traditional fully supervised deep learning approaches, which require pixel-level annotation of every inclusion body, necrotic focus, and inflammatory cell, are prohibitively labor-intensive and impractical for the scale required in veterinary diagnostics [6, 18]. This has driven the field toward weakly supervised learning paradigms, which can operate using only slide-level diagnostic labels (e.g., "positive for African Swine Fever Virus" versus "negative") without requiring pathologists to manually outline every lesion. These methods, built upon the framework of Multiple Instance Learning (MIL), treat each WSI as a "bag" of smaller image patches (instances). The model learns to identify which instances are most predictive of the slide-level label, effectively performing automated attention on regions of diagnostic significance [10, 15, 41].

Architectural Innovations for Viral Lesion Detection

The evolution of neural network architectures has been instrumental in advancing viral lesion classification. Convolutional Neural Networks (CNNs), particularly variants like ResNet and EfficientNet, have demonstrated robust performance in patch-level classification tasks, successfully differentiating between normal tissue, benign lesions, and viral-induced neoplasia in contexts such as Marek's Disease Virus-induced lymphomas in poultry [9, 13]. However, CNNs have inherent limitations in capturing the long-range spatial dependencies and hierarchical tissue organization that are crucial for recognizing certain viral lesions. For example, the progression of Feline Coronavirus and FIP-associated pyogranulomatous inflammation involves a complex interplay of vasculitis, perivascular cuffing, and serosal involvement that spans large areas of tissue, a context that pure CNNs struggle to model effectively.

Transformer-based architectures, specifically Vision Transformers (ViTs), have emerged as a powerful alternative by employing self-attention mechanisms that can weigh the relevance of every image patch relative to every other patch, effectively modeling global tissue context [19, 39]. This capability is particularly valuable for viral lesions that are multifocal or exhibit zonal distributions. Studies comparing CNNs and ViTs in histopathology have consistently shown the superiority of transformers for tasks requiring spatial reasoning across large tissue expanses, such as grading the extent of hepatic necrosis in Rift Valley Fever Virus infection or quantifying the severity of enteritis in Porcine Epidemic Diarrhea Virus [39].

More recent innovations, such as the Mamba architecture, address the computational inefficiency of standard transformers by enabling selective state-space modeling. The HCSMIL framework, for instance, demonstrated a marked improvement in recognizing small, complexly distributed lesions-a scenario directly analogous to identifying early viral inclusion bodies or focal necrotic hepatocytes in Infectious Salmon Anemia Virus infections-achieving an 84% recognition rate for small lesions compared to 70.59% for traditional MIL approaches [39]. This underscores the critical role of architectural choice in capturing the specific spatial topologies of viral cytopathology.

Feature Extraction and Representation Learning

The efficacy of any AI model for viral lesion classification is fundamentally dependent on the quality of the features it learns to extract. Traditional approaches relied on handcrafted features describing tissue composition, nuclear morphology, and texture. While useful for certain applications-such as quantifying the proportion of seminiferous tubule damage in Channel Catfish Virus infections using tools like ilastik and Fiji [34]-these methods lack the adaptability to generalize across the vast morphological spectrum of viral diseases.

Contemporary approaches leverage self-supervised learning (SSL) to learn rich, transferable feature representations from unlabeled WSIs. Methods like contrastive learning train models to recognize that different views (e.g., differently stained or rotated patches) of the same tissue region should have similar feature embeddings, while patches from different regions should be distinct [2]. This is particularly valuable in veterinary virology, where datasets are often imbalanced-with abundant negative slides and sparse positive cases-and where stain variability across laboratories and species can confound models. SSL-based pre-training on large, diverse WSI cohorts has been shown to produce feature extractors that are remarkably robust to such domain shifts, improving classification of lesions caused by Porcine Reproductive and Respiratory Syndrome Virus or Bovine Viral Diarrhea Virus across different staining protocols [7, 24].

A particularly exciting frontier is the integration of textual pathological knowledge to guide visual representation learning. The PathTree framework, for example, represents diagnostic categories as hierarchical, text-based pathological descriptions-such as "hepatocellular necrosis with Councilman bodies" for Rabies Lyssavirus or "intraepidermal vesicle with ballooning degeneration" for Foot-and-Mouth Disease Virus. By aligning visual features with these text prototypes through cross-modal contrastive learning, the model can leverage expert domain knowledge to improve classification accuracy and provide a degree of explainability that is critical for clinical trust [8, 12].

Segmentation and Quantification: Moving Beyond Classification

While slide-level classification (e.g., "viral hepatitis present") is valuable, the true power of AI in viral lesion analysis lies in its ability to perform precise, reproducible segmentation and quantification of specific histological features. This capability is essential for understanding disease pathogenesis, staging infection severity, and monitoring therapeutic interventions. Deep learning-based segmentation models can delineate the boundaries of viral inclusion bodies, map the extent of necrotic foci, and quantify the density and spatial distribution of infiltrating immune cells [14, 44].

For instance, in the context of Ranaviruses in Amphibians, which cause systemic hemorrhagic disease, AI models can segment and count intracytoplasmic inclusion bodies across dozens of organs, providing a quantitative viral load proxy directly from H&E-stained slides. Similarly, in White Spot Syndrome Virus infections in shrimp, automated segmentation of pathognomonic intranuclear inclusions in ectodermal and mesodermal tissues enables high-throughput screening of aquaculture populations [44]. The incremental relationship-guided segmentation (IRS) framework addresses a further challenge: the need for models to incorporate new disease phenotypes or lesion types over time without catastrophic forgetting. This "continual learning" paradigm is ideally suited for the evolving landscape of viral diseases, where emerging pathogens like Tilapia Lake Virus or novel strains of Infectious Bursal Disease Virus require model adaptation without full retraining [14].

Multimodal Integration and Predictive Modeling

The most powerful AI systems for viral lesion classification are those that transcend the WSI itself, integrating histomorphological data with other modalities to create a holistic, predictive model of disease. This multimodal approach mirrors the diagnostic reasoning of the expert pathologist, who synthesizes histological findings with clinical history, gross pathology, serology, and molecular diagnostics [1, 30, 42]. For example, a model predicting the likelihood of histologic fibrosis reversal in Bovine Viral Diarrhea Virus persistently infected calves could integrate WSI features with clinical parameters (age, viral load, liver enzyme levels) to produce a more accurate prognosis [27]. This fusion of imaging and structured data has been shown to outperform models based on either modality alone, achieving area under the curve (AUC) improvements of 0.694 versus 0.588 for clinical-only models in predicting fibrosis outcome [27].

The potential for AI to act as a digital biomarker is particularly compelling for viral lesions where histological grade is a critical determinant of clinical outcome. In Marek's Disease Virus-induced lymphomas, for instance, the presence and number of transformed T-cells within visceral organs is directly correlated with mortality. AI models that can not only detect the presence of lymphoma but also predict disease progression based on subtle architectural features within the WSI-features that may be imperceptible to the human eye-could revolutionize flock management and vaccine efficacy trials [40]. Work in human oncology has demonstrated that AI models can extract prognostic information from H&E-stained slides alone, predicting disease recurrence in low-risk breast cancers with an agreement of 66% compared to the gold-standard Oncotype DX genomic test [40]. The translation of such "histogenomic" approaches to veterinary virology is an active area of research, with potential applications in predicting the clinical trajectory of Feline Coronavirus and FIP or Equine Herpesvirus 1-induced myeloencephalopathy.

Benchmarking, Validation, and the Path to Clinical Deployment

The clinical deployment of AI models for viral lesion classification demands rigorous, multi-institutional validation to ensure generalizability across diverse populations, tissue types, and scanning platforms [9, 33]. Early studies comparing AI performance to pathologist consensus, such as those on urothelial neoplasms using 12,500 WSIs from five institutions, have demonstrated that well-trained models-particularly EfficientNet-B6-can achieve AUCs of 0.983 with accuracy of 0.913, approaching expert-level performance [9]. However, these benchmarks are typically established in controlled settings with standardized slide preparation and well-defined lesion categories.

The translation to real-world veterinary practice presents additional challenges. Viral lesions often occur in tissues with complex backgrounds-such as the lymphoid tissue of the bursa in Infectious Bursal Disease Virus or the gill epithelium in Infectious Salmon Anemia Virus-where pre-existing inflammation, artifact, or normal structural variation can confound model predictions. The variability in tissue processing and scanning protocols across veterinary diagnostic laboratories introduces domain shift, reducing model performance [23]. Solutions to this include stain normalization techniques, domain adversarial training, and the construction of large, diverse training cohorts that capture the full spectrum of biological and technical variation [12, 23].

Furthermore, the interpretability of AI models is paramount for clinical acceptance. A model that flags a WSI as "positive for West Nile Virus in Birds" is of limited value if it cannot explain which histological features drove that decision. Attention-based MIL models address this by generating heatmaps that highlight the regions of the slide that were most influential in the classification, effectively guiding the pathologist's eye to potential lesions [7, 15]. Recent advances in concept-based explanation methods can link these attention maps to specific histopathological entities, such as "lymphocytic cuffing of cerebral vessels" for West Nile Virus or "myocardial necrosis with lymphohistiocytic infiltration" for West Nile Virus in Birds [8]. This not only builds trust but also serves as a valuable educational tool for trainees and a quality control mechanism for experienced pathologists [38, 43].

The ultimate integration of AI into the clinical workflow will likely take the form of a "digital assistant" that triages WSIs, flags suspicious regions for review, and provides quantitative metrics to support diagnostic decision-making. In the context of large-scale surveillance for foreign animal diseases-such as African Swine Fever Virus or Classical Swine Fever Virus-a first-pass AI screening system could dramatically reduce the workload on expert pathologists, allowing them to focus their efforts on the most challenging or high-priority cases [24]. Studies on automated detection of BVDV antigen in ear-notch specimens have demonstrated that an AI-assisted workflow can achieve 97.4% sensitivity while significantly reducing personnel time, with the model effectively serving as a pre-screening tool [24]. Similar approaches are being explored for the rapid identification of White Spot Syndrome Virus in shrimp aquaculture and Ranaviruses in Amphibians in wildlife surveillance programs [44]. The path forward requires continued collaboration between veterinary pathologists, computer scientists, and regulatory bodies to establish standards for model validation, deployment, and ongoing performance monitoring, ensuring that these powerful tools enhance, rather than supplant, expert diagnostic judgment.

Challenges in Standardization, Validation, and Regulatory Considerations

The integration of digital pathology and whole-slide imaging (WSI) into the diagnostic workflow for viral lesions represents a paradigm shift in veterinary medicine, yet its widespread adoption is contingent upon surmounting formidable barriers in standardization, analytical validation, and regulatory oversight. Unlike static histopathologic assessments of neoplasia, viral lesions present unique complexities: they are often multifocal, temporally dynamic, and intimately linked to host immune responses that can obscure or mimic cytopathic effects. The stakes are exceptionally high, as misdiagnosis of notifiable pathogens-such as African Swine Fever Virus, Avian Influenza Virus, or Rabies Lyssavirus-can have catastrophic consequences for animal health, trade, and public safety. This section critically examines the principal challenges that must be addressed to transform WSI-based viral lesion analysis from a promising research tool into a validated, regulatory-compliant diagnostic modality.

Pre-Analytical and Analytical Standardization: The Foundation of Reproducibility

The journey of a tissue section from the necropsy table to the digital slide is fraught with variability that directly undermines the reliability of AI-driven interpretation. Pre-analytical factors-including tissue fixation protocols, section thickness, staining batch effects, and slide mounting-introduce systematic biases that are amplified when WSI is subjected to computational analysis. In the context of viral lesions, these issues are particularly acute because viral cytopathology is exquisitely sensitive to tissue handling. For instance, autolysis can rapidly degrade viral inclusion bodies or syncytia, rendering them undetectable by both pathologists and algorithms [21]. Studies evaluating melanocytic lesions have demonstrated that even expert pathologists exhibit significant diagnostic variability when interpreting WSIs compared to glass slides, with z-stack scanning offering only marginal improvements in diagnostic accuracy for challenging entities such as nevoid melanoma [16]. This finding is directly translatable to viral lesions, where subtle intranuclear or intracytoplasmic inclusions (e.g., those seen in Canine Distemper Virus or Bovine Herpesvirus 1 infection) may be missed without the ability to fine-focus through multiple planes.

The issue of inter-institutional scanning variability is starkly illustrated by a multicenter study of WSIs for melanoma diagnosis in children, adolescents, and young adults, where 46.3% of 311 WSIs exhibited technical defects-70.1% of which were analytical (scanning-related) and 29.9% pre-analytical [23]. This study documented significant inter-center variability in file sizes, reflecting non-uniform scanning protocols and tissue processing standards. For veterinary viral diagnostics, where samples may originate from field necropsies, slaughterhouses, or wildlife surveillance programs, such variability is the norm rather than the exception. The lack of standardized scanning parameters-including resolution (e.g., 20× vs. 40×), compression algorithms, and color calibration-poses a fundamental challenge for any downstream AI model trained on images from a single scanner or institution [1, 2]. The use of digital slide middleware, as proposed by some groups, offers a potential solution by providing real-time quality control and rescaling, but this technology remains nascent in veterinary applications [17].

Analytical Validation: Establishing Ground Truth for Viral Lesions

The validation of AI models for WSI-based viral lesion detection is hampered by the absence of a universally accepted reference standard. In human digital pathology, ground truth is typically established by expert panel consensus, yet even among board-certified dermatopathologists, concordance rates for melanocytic lesions range from 75% to 90%, with nevoid melanoma showing alarming variability of 10% to 80% [16]. For veterinary viral infections, the situation is even more complex. The histopathologic diagnosis of viral disease often relies on a constellation of findings-including cell tropism, lesion distribution, inflammatory infiltrate composition, and the presence of pathognomonic inclusion bodies-that are subject to considerable inter-observer variability even among experienced veterinary pathologists [4, 24]. This was elegantly demonstrated in a study of Bovine Viral Diarrhea Virus antigen detection in ear-notch specimens using digital image analysis, where the automated system achieved 97.4% sensitivity and 89.4% specificity compared to manual review [24]. While impressive, the 10.6% false-positive rate underscores the challenge of establishing a reliable gold standard when the reference method itself (manual immunohistochemistry interpretation) is known to have imperfect reproducibility.

Furthermore, the validation of AI models for viral lesions must account for the biological spectrum of infection. Unlike neoplastic lesions, which are generally static at the time of biopsy, viral lesions evolve over the course of infection. Early-stage lesions may show only subtle cytopathic effects (e.g., cell swelling, minimal apoptosis), while fulminant infections are characterized by extensive necrosis, hemorrhage, and secondary bacterial invasion that confound algorithmic feature extraction [4, 8]. The Wesselsbron virus-induced hepatitis study demonstrated that machine learning-driven digital histopathology could reliably quantify lymphohistiocytic infiltration and hepatocyte proliferation indices that correlated with viral load and serum markers of hepatic injury [4]. However, this study involved experimentally infected animals under controlled conditions, which do not reflect the heterogeneity of field cases where coinfections, nutritional status, and host genetics introduce additional variance. The challenge of detecting subtle or rare lesions-a hallmark of many chronic viral infections such as Maedi-Visna Virus or Feline Leukemia Virus-is a recognized limitation of current weakly supervised learning approaches, which often fail to capture diagnostically critical but spatially restricted features [7, 39]. Indeed, a multi-directional context modeling framework (HCSMIL) achieved only 84% small-lesion recognition rate, highlighting the persistent difficulty in detecting the very features that define early or atypical viral pathology [39].

Computational Challenges: Imbalanced Data and Domain Shift

The application of deep learning to WSI analysis for viral lesions is further complicated by the intrinsic characteristics of the training data. Viral diseases often have a low prevalence in surveillance populations, leading to severe class imbalance in which negative (non-lesional) tiles overwhelmingly outnumber positive tiles. This imbalance is a well-recognized challenge in digital pathology, and contrastive learning strategies have been proposed to improve minority class recognition [2]. However, these methods are typically validated on curated datasets with distinct lesion boundaries (e.g., tumor margins), whereas viral lesions frequently exhibit ill-defined, zonal, or diffuse patterns of infection. For example, the neurotropism of Eastern Equine Encephalitis Virus in Birds or the hepatotropism of Duck Hepatitis A Virus produces lesions that are distributed across large tissue areas with variable severity, making tile-level annotation ambiguous [6, 15]. Multiple-instance learning (MIL) frameworks, which operate under the assumption that only a subset of patches (instances) within a WSI (bag) are positive, have emerged as a popular solution [15, 37]. Yet, MIL models trained on slide-level labels alone can be misled by confounding factors such as staining artifacts, tissue folding, or concurrent non-viral pathology, leading to spurious correlations [18, 41].

Domain shift represents another formidable obstacle. The appearance of viral lesions on WSI is heavily influenced by the scanner characteristics, staining protocol (e.g., hematoxylin and eosin vs. immunohistochemistry), and tissue processing pipeline used at each institution. Models trained on data from a single center often fail to generalize to external datasets, as evidenced by studies showing significant performance degradation when algorithms are tested on out-of-distribution cohorts [5, 9, 20]. This is particularly problematic for veterinary viral diagnostics, where samples may originate from diverse geographic regions with distinct histology laboratory practices. A weakly supervised model for classifying fibro-osseous lesions of the jaw achieved an area under the curve of 0.86 in internal testing but was noted to require further validation in more diverse cohorts to substantiate generalizability [20]. Similarly, a deep learning model for bladder cancer classification showed high performance (AUC 0.983) across five institutions but relied on pre-processing steps such as stain normalization, which adds computational overhead and may not fully correct for batch effects [9]. For viral lesions, where the histomorphology itself may be subtle and overshadowed by staining variability, domain adaptation techniques are not merely beneficial but essential.

Regulatory Hurdles and the Path to Clinical Adoption

The regulatory landscape for AI-enabled digital pathology in veterinary medicine is fragmented and ill-equipped to handle the complexities of viral lesion analysis. In contrast to human medicine, where agencies such as the U.S. Food and Drug Administration (FDA) have established frameworks for software as a medical device (SaMD), veterinary diagnostic devices are often subject to less stringent oversight, varying by jurisdiction. The World Organisation for Animal Health (WOAH) provides guidelines for diagnostic test validation, but these were designed for traditional assays (e.g., PCR, ELISA) and do not adequately address the unique considerations of computational pathology [25]. A major regulatory challenge lies in defining the intended use of the AI system. Is it intended for primary diagnosis, triage, or as a second-reader tool? Each use case carries different risk profiles and validation requirements. For high-consequence pathogens such as Foot-and-Mouth Disease Virus or Classical Swine Fever Virus, a false-negative result from an AI-driven screening tool could delay outbreak detection and lead to catastrophic spread. Conversely, false-positive results could trigger unnecessary stamping-out policies with severe economic repercussions.

The issue of interpretability compounds regulatory concerns. Deep learning models, particularly vision transformers and attention-based MIL frameworks, are often considered "black boxes" that provide probability scores without transparent reasoning [1, 30]. For veterinary pathologists to trust and adopt these tools, the models must provide explainable outputs-such as heatmaps highlighting the regions of the WSI that most strongly influenced the prediction [7, 8]. While attention-based mechanisms offer a degree of interpretability, they can also be misleading, focusing on staining artifacts or normal tissue structures rather than true viral cytopathology [12, 15]. The development of pathology-attention frameworks that incorporate expert-defined text prototypes or hierarchical knowledge graphs represents a promising step toward integrating domain expertise with computational analysis [8, 12]. However, these approaches remain experimental and have not been validated in the high-stakes context of veterinary viral diagnostics.

Data Governance, Privacy, and the Need for Multi-Institutional Collaboration

The development of robust, generalizable AI models for viral lesion analysis requires access to large, diverse, and high-quality datasets that are currently lacking in veterinary medicine. Unlike human digital pathology, where repositories such as The Cancer Genome Atlas (TCGA) provide thousands of curated WSIs, veterinary repositories are fragmented, pathogen-specific, and often limited to single-institution collections [21, 44]. The rarity of certain viral diseases (e.g., Abalone Herpesvirus or Decapod Iridescent Virus 1) makes data collection logistically challenging, and the samples that do exist are often archived in formats (e.g., glass slides) that require costly digitization [23]. Privacy and data-sharing concerns further impede collaboration. While animal health data may not carry the same patient privacy implications as human medical records, proprietary concerns of pharmaceutical companies, competitive interests of diagnostic laboratories, and national security issues surrounding select agents (e.g., Rift Valley Fever Virus) can restrict data access [26, 45].

The establishment of centralized digital pathology platforms, analogous to the Picture Archiving and Communication Systems (PACS) used in human radiology, is a prerequisite for overcoming these barriers [17, 30]. These platforms must incorporate robust metadata standards that capture pre-analytical variables (e.g., fixation time, embedding protocol, scanner model), clinical metadata (e.g., species, age, vaccination status, PCR results), and pathologist annotations. Without such standardization, the "curse of dimensionality" inherent in WSI analysis is compounded by unmeasured confounding, leading to models that are brittle and non-reproducible [1, 21]. Federated learning-a technique that trains models across multiple institutions without centralizing raw data-offers a promising solution for preserving data sovereignty while enabling model generalizability [46]. However, this approach requires sophisticated information technology infrastructure and consensus on model architecture, hyperparameters, and validation protocols that are challenging to achieve across diverse veterinary settings.

The Pathologist's Role in an AI-Augmented Future

Ultimately, the successful integration of WSI-based AI for viral lesion analysis hinges on recognizing that the technology is a decision-support tool, not a replacement for the veterinary pathologist. The toxicologic pathology community has provided cautious perspectives on this transition, noting that while digital platforms have great potential, they are not yet ready for complete replacement of glass slides in safety assessments [25]. This caution is equally or more applicable to viral diagnostics, where the consequences of error are severe and the lesions are inherently complex. As demonstrated in studies of cervical biopsies and melanocytic lesions, AI can improve diagnostic accuracy and interobserver agreement, but it also introduces new failure modes, particularly for rare or morphologically deceptive entities [16, 22]. The imperative for rigorous, multi-institutional validation, adherence to regulatory standards set forth by agencies such as the WOAH and FAO, and the maintenance of pathologist oversight cannot be overstated. Only through a framework that couples technical rigor with clinical expertise can digital pathology realize its potential to revolutionize the diagnosis of viral lesions in veterinary medicine.

Future Directions: Multimodal Data Fusion and Precision Virology

The trajectory of digital pathology for viral lesions is poised for a transformative leap beyond the mere digitization of histomorphology. The next frontier, which we term precision virology, is predicated on the synergistic integration of whole-slide imaging (WSI) data with a panoply of orthogonal data streams-genomic, transcriptomic, proteomic, metabolomic, and clinical-to construct a holistic, mechanistically informed model of viral pathogenesis. This paradigm shift, from descriptive histopathology to a predictive, multi-dimensional science, is not merely an incremental advance but a fundamental re-conceptualization of how we diagnose, prognosticate, and ultimately manage viral diseases in both clinical and veterinary settings [1, 30, 45]. The core tenet is that the tissue lesion captured in a WSI is the final common pathway of a complex interplay between viral virulence factors, host genetic susceptibility, and dynamic immune responses. To unravel this complexity, we must move beyond univariate analyses and embrace the power of multimodal data fusion.

The Paradigm Shift: From Morphology to Multidimensional Pathogenesis

Traditional virology has often relied on a reductionist approach, correlating a single viral load measurement or a crude histopathological score with clinical outcome. This approach fails to capture the spatial heterogeneity and temporal dynamics of infection. For instance, the histopathological signature of Avian Influenza Virus in the respiratory tract-ranging from mild catarrhal inflammation to severe necrotizing bronchiolitis-is not solely a function of viral cytopathicity but is profoundly shaped by the host's local and systemic immune microenvironment. A multimodal framework would integrate the WSI-derived spatial distribution of viral antigen (e.g., via immunohistochemistry) with transcriptomic data from the same tissue block, revealing the specific immune cell subsets (e.g., cytotoxic T lymphocytes, macrophages, regulatory T cells) that are recruited and their activation states [3, 28]. This approach has been elegantly demonstrated in a study of Wesselsbron virus-induced hepatitis, where machine learning-driven digital histopathology quantified T cell and histiocyte densities, which were then correlated with viral load and serum biomarkers of liver injury, providing a multidimensional view of disease severity that no single parameter could achieve [4].

The application of such integrated analyses to aquatic viral diseases is particularly compelling. Consider White Spot Syndrome Virus in shrimp, where the characteristic intranuclear inclusion bodies are pathognomonic but provide limited insight into the host's antiviral response. By fusing WSI data with transcriptomic profiles from the same infected tissue, we could map the spatial expression of key immune effectors like prophenoloxidase and antimicrobial peptides relative to the viral inclusion bodies, thereby identifying tissue microenvironments that are permissive or restrictive to viral replication. Similarly, for Infectious Salmon Anemia Virus, the fusion of histopathological lesion severity in the gills and kidney with viral genotyping and host transcriptomic signatures could differentiate between high and low pathogenicity strains and predict the likelihood of clinical disease progression [1, 45]. This represents a move from a purely diagnostic exercise to a prognostic and mechanistic one.

Architectures for Multimodal Fusion in Viral Pathology

The computational architectures required for effective multimodal fusion are rapidly maturing. Early efforts often employed a "late fusion" strategy, where separate models were trained on histology, genomics, and clinical data, and their predictions were combined at the decision level [27]. While straightforward, this approach fails to capture the complex, non-linear interactions between modalities. More sophisticated "early fusion" or "intermediate fusion" models, such as the pathology-attention multi-instance learning (PAT-MIL) framework, are now emerging. PAT-MIL integrates visual features from WSIs with expert-defined text prototypes (e.g., pathological descriptions of lesion types) using dynamic attention mechanisms, thereby guiding the model to focus on histomorphological features that are semantically aligned with clinical knowledge [12]. This concept can be extended to virology by incorporating text prototypes derived from viral pathogenesis literature-for example, descriptions of syncytia formation, intranuclear inclusions, or apoptotic bodies-to create a knowledge-guided AI that is not merely a "black box" but an interpretable assistant.

A particularly promising direction is the use of contrastive learning and vision-language models to align histopathological images with molecular data. The MultiXpert framework, developed for chest X-ray diagnosis, demonstrates how a dual-stream architecture can synergistically enhance image and text representations through cross-modal alignment [47]. In a virology context, such a model could be trained to align a WSI patch showing a characteristic viral lesion (e.g., the hepatocellular necrosis and Councilman bodies of Rift Valley Fever Virus) with the corresponding viral RNA sequence or a quantitative PCR result. This would enable zero-shot or few-shot classification of novel viral lesions based on their histomorphological similarity to previously characterized pathogens, a capability of immense value for emerging infectious disease surveillance [8, 47]. The challenge of imbalanced datasets, where certain viral lesions are rare (e.g., Tilapia Lake Virus in certain regions), can be addressed through self-supervised and contrastive learning strategies that learn robust representations from abundant unlabeled data before fine-tuning on scarce labeled examples [2, 6].

Precision Virology: Stratifying Host Response and Predicting Outcome

The ultimate goal of multimodal data fusion is to enable precision virology-the stratification of infected individuals into distinct risk categories based on a comprehensive assessment of host-pathogen interaction. This goes far beyond simple viral detection. For example, in African Swine Fever Virus infection, the clinical outcome ranges from peracute death to chronic, asymptomatic infection. A multimodal model that integrates WSI features (e.g., degree of lymphoid depletion, macrophage infiltration, and fibrinoid necrosis) with host genomic markers of immune response (e.g., polymorphisms in MHC or interferon genes) and viral genotyping could predict, at the time of diagnosis, whether an individual pig is likely to become a superspreader or to mount an effective immune response [1, 42]. This would revolutionize outbreak management, allowing for targeted culling, quarantine, or even early therapeutic intervention.

The integration of digital pathology with advanced molecular diagnostics, such as in situ hybridization (ISH) and multiplexed immunohistochemistry (IHC), is a critical enabler. While standard IHC for a single viral antigen provides limited information, multiplexed IHC panels that simultaneously detect viral proteins, host immune cell markers (CD3, CD20, Iba-1), and functional markers (Ki67, granzyme B) can generate a rich spatial map of the immune microenvironment [3, 4]. When combined with WSI analysis, these data can be used to construct "digital twin" models of the infected tissue-a computational representation that can be perturbed in silico to simulate the effects of antiviral drugs or immunomodulatory therapies [1]. For instance, a digital twin of a Porcine Reproductive and Respiratory Syndrome Virus-infected lung could predict whether a specific anti-inflammatory therapy would reduce macrophage infiltration and fibrosis without impairing viral clearance.

Technical Hurdles and the Path to Clinical Deployment

Despite its immense promise, the path to clinical deployment of multimodal fusion in virology is fraught with technical and logistical hurdles. A primary challenge is data heterogeneity and standardization. WSIs are generated using different scanners with varying resolution, color profiles, and compression algorithms, while genomic and transcriptomic data are produced using diverse platforms and bioinformatics pipelines [23, 25]. The lack of standardized, interoperable data formats creates significant barriers to model development and multi-institutional validation. The development of robust middleware and quality control pipelines, analogous to those used in clinical chemistry, is essential to ensure that data fed into multimodal models are of consistent quality [17]. Furthermore, the computational cost of processing and fusing massive datasets-a single WSI can be several gigabytes, and a whole-transcriptome dataset even larger-requires scalable cloud-based or high-performance computing infrastructure. Platforms like ComPRePS, which automate WSI ingestion and feature extraction in a cloud environment, represent a step in the right direction, but they must be extended to handle non-image data [35].

Another critical issue is model interpretability. For a pathologist or virologist to trust a multimodal model's prediction, they must understand why the model arrived at that conclusion. Attention-based mechanisms, such as those used in the HCSMIL framework, can highlight the specific regions of a WSI that were most influential in the model's decision, while gradient-weighted class activation mapping (Grad-CAM) can overlay heatmaps onto the tissue to show where the model is "looking" [27, 39]. However, interpreting the interaction between modalities-for example, why a specific histopathological feature combined with a particular gene expression signature predicts a poor outcome-remains a significant challenge. The development of explainable AI (XAI) methods tailored for multimodal medical data is a high-priority research area [26, 46]. Finally, rigorous, multi-institutional validation is paramount. Models trained on data from a single laboratory or geographic region may fail to generalize due to differences in host genetics, viral strains, or tissue processing protocols. The success of models like PathologAI in discriminating treatment-related from spontaneous necrosis in rat livers, and the multi-institutional validation of urothelial neoplasm classifiers, provide a template for the kind of large-scale, collaborative validation studies that will be required to bring multimodal virology models into routine diagnostic and regulatory use [6, 7, 9]. The field must now pivot from proof-of-concept studies to the systematic, evidence-based implementation of these powerful tools.

References

[1] Huynh C. Artificial intelligence and the evolution of pathology: Scaling from digital diagnostics to multimodal precision medicine. Intelligence-Based Medicine. 2026. DOI: https://doi.org/10.1016/j.ibmed.2026.100386

[2] Roy S, Jadhav R, Meena T. Contrastive learning strategies for better image classification with imbalanced datasets . Applied Soft Computing. 2026. DOI: https://doi.org/10.1016/j.asoc.2025.114450

[3] Heide Vvd, McArdle S, Nelson MS, Cerosaletti K, Gnjatic S, Mikulski Z, et al.. Integrated histopathology of the human pancreas throughout stages of type 1 diabetes progression. Nature Communications. 2026. DOI: https://doi.org/10.1038/s41467-026-68610-1

[4] Grau-Roma L, Brot Sd, Zimoch M, Clerc L, Donzé N, Liniger M, et al.. Wesselsbron Virus‐Induced Hepatitis in Ewes and Lambs Unraveled Through Machine Learning‐Driven Digital Histopathology. Transboundary and Emerging Diseases. 2026. DOI: https://doi.org/10.1155/tbed/7912840

[5] Tauqeer A, Asif A, Sadeghi-Naini A. Detection, localization, and staging of breast cancer lymph node metastasis in digital pathology whole slide images using selective neighborhood attention-based deep learning. Scientific Reports. 2025. DOI: https://doi.org/10.1038/s41598-025-21787-9

[6] Bussola N, Xu J, Wu L, Gorini L, Zhang Y, Furlanello C, et al.. A Weakly Supervised Deep Learning Framework for Whole Slide Classification to Facilitate Digital Pathology in Animal Study. Chemical Research in Toxicology. 2023. DOI: https://doi.org/10.1021/acs.chemrestox.3c00058

[7] Zehnder P, Feng J, Nguyen T, Shen P, Sullivan R, Fuji RN, et al.. Diagnostic classification in toxicologic pathology using attention-guided weak supervision and whole slide image features: a pilot study in rat livers. Scientific Reports. 2025. DOI: https://doi.org/10.1038/s41598-025-86615-6

[8] Li J, Sun Q, Yan R, Wang Y, Fu Y, Wei Y, et al.. Diagnostic text-guided representation learning in hierarchical classification for pathological whole slide image. Medical Image Anal.. 2025. DOI: https://doi.org/10.1016/j.media.2025.103894

[9] Park JY, Kim J, Kim YJ, Kim SH, An CS, Kim KG, et al.. Multi-institutional validation of AI models for classifying urothelial neoplasms in digital pathology. Scientific Reports. 2025. DOI: https://doi.org/10.1038/s41598-025-21096-1

[10] Aswolinskiy W, Post RSvd, Campora M, Baronchelli C, Ardighieri L, Vatrano S, et al.. Attention-based whole-slide image compression achieves pathologist-level pre-screening of multi-organ routine histopathology biopsies. medRxiv. 2024. DOI: https://doi.org/10.1101/2024.12.17.24319180

[11] Wang K, Liu R, Chen Y, Wang Y, Gao Y, Qiu Y, et al.. Accurate diagnosis achieved via super-resolution whole slide images by pathologists and artificial intelligence. medRxiv. 2024. DOI: https://doi.org/10.1101/2024.07.05.24310022

[12] Fu F, Zhang X, Wang Z, Xie L, Fu M, Peng J, et al.. A pathology-attention multi-instance learning framework for multimodal classification of colorectal lesions. Frontiers in Pharmacology. 2025. DOI: https://doi.org/10.3389/fphar.2025.1592950

[13] Soldatov S, Pashkov D, Guda S, Karnaukhov N, Guda A, Soldatov AV. Deep Learning Classification of Colorectal Lesions Based on Whole Slide Images. Algorithms. 2022. DOI: https://doi.org/10.3390/a15110398

[14] Deng R, Zhu J, Xiong J, Cui C, Yao T, Guo J, et al.. IRS: Incremental Relationship-guided Segmentation for Digital Pathology. arXiv.org. 2025. DOI: https://doi.org/10.48550/arXiv.2505.22855

[15] Zhang X, Liu C, Zhu H, Wang T, Du Z, Ding W. A universal multiple instance learning framework for whole slide image analysis. Comput. Biol. Medicine. 2024. DOI: https://doi.org/10.1016/j.compbiomed.2024.108714

[16] Sturm B, Creytens D, Cook M, Smits J, Dijk MVv, Eijken E, et al.. Validation of Whole-slide Digitally Imaged Melanocytic Lesions: Does Z-Stack Scanning Improve Diagnostic Accuracy?. Journal of Pathology Informatics. 2019. DOI: https://doi.org/10.4103/jpi.jpi_46_18

[17] Smith G. 61 Digital Slide Middleware Increases the Quality of a Digital Primary Diagnosis Pipeline. American Journal of Clinical Pathology. 2025. DOI: https://doi.org/10.1093/ajcp/aqaf121.488

[18] Chen C, Chen C, Yu W, Chen S, Chang Y, Hsu T, et al.. An annotation-free whole-slide training approach to pathological classification of lung cancer types using deep learning. Nature Communications. 2021. DOI: https://doi.org/10.1038/s41467-021-21467-y

[19] Liu-Swetz Y, Niksic S, Seethala R, Shasteen AM, Foran D, Bilodeau EA. AI-driven prediction of progression to oral squamous cell carcinoma using a multiresolution pathology model. npj Digital Medicine. 2025. DOI: https://doi.org/10.1038/s41746-025-02014-1

[20] Zhang AB, Li P, Xue J, Zhang J, You Z, Ge S, et al.. Deep Learning on Histology Images for Differentiating Fibro-osseous Lesions of the Jaw.. Journal of dentistry research. 2026. DOI: https://doi.org/10.1177/00220345261415888

[21] Almeida ALTd, Santos ADd, Barreto-Vieira DF. Cracking the Code: Computational Image Analysis Tools for Histopathological and Morphometric Insights. Journal of Imaging. 2026. DOI: https://doi.org/10.3390/jimaging12040173

[22] Andreassen AK, Mortensen E, Stenbro R, Sørensen Ø, Sørbye SW. Digital Pathology with AI for Cervical Biopsies: Diagnostic Accuracy at the CIN2+ Threshold. Cancers. 2025. DOI: https://doi.org/10.3390/cancers17233808

[23] Intergroup IM. 11 | Challenges in developing a digital pathology consultation network: insights from the melcaya experience. Dermatology Reports. 2025. DOI: https://doi.org/10.4081/dr.2025.10709

[24] Lin SJ, Magstadt DR, Derscheid R, Burrough ER. Using HALO digital image analysis for automated detection of bovine viral diarrhea virus antigen in ear-notch specimens. Journal of Veterinary Diagnostic Investigation. 2025. DOI: https://doi.org/10.1177/10406387241307643

[25] Schafer K, Rao DB. Toxicologic Pathology Forum: Opinion on Digital Primary Read and Peer Review - Are We There Yet?. Toxicologic pathology (Print). 2025. DOI: https://doi.org/10.1177/01926233241303911

[26] Alam R, Raiyana J. Pathology AI: Safe Slide, Sample, and Medical Report Analysis in Cancer Care. International Journal of Medical and Health Research. 2025. DOI: https://doi.org/10.61424/ijmhr.v3i4.630

[27] Han W, Cheng D, He Q, Wang S, Gong S, Chen Y, et al.. Deep learning-based multimodal model for predicting on-treatment histological outcomes in chronic hepatitis B-associated advanced liver fibrosis. World Journal of Gastroenterology. 2026. DOI: https://doi.org/10.3748/wjg.v32.i15.116679

[28] Zhang Y, Gupta R, Saltz J, Kurç T, Kaczmarzyk JR, Bremer E, et al.. Abstract 5315: Mapping tumor infiltrating lymphocytes in whole prostatectomy specimens to visualize and quantify the immune microenvironment of prostate cancer. Cancer Research. 2025. DOI: https://doi.org/10.1158/1538-7445.am2025-5315

[29] Rhoads D, Habib-Bein N, Hariri RS, Hartman D, Monaco SE, Lesniak A, et al.. Comparison of the diagnostic utility of digital pathology systems for telemicrobiology. Journal of Pathology Informatics. 2016. DOI: https://doi.org/10.4103/2153-3539.177687

[30] Ramalhete L, Araújo R, Vieira MB, Vigia E, Calado CRC, Ferreira A. Rejection-Focused Precision Medicine in Kidney Transplantation: Biology, Biomarkers, and Artificial Intelligence. Life. 2026. DOI: https://doi.org/10.3390/life16040674

[31] Reyes C, Ikpatt O, Nadji M, Cote R. Intra-observer reproducibility of whole slide imaging for the primary diagnosis of breast needle biopsies. Journal of Pathology Informatics. 2014. DOI: https://doi.org/10.4103/2153-3539.127814

[32] Ibrahim A, Lashen AG, Toss M, Mihai R, Rakha E. Assessment of mitotic activity in breast cancer: revisited in the digital pathology era. Journal of Clinical Pathology. 2021. DOI: https://doi.org/10.1136/jclinpath-2021-207742

[33] Malarkey D, Willson GA, Willson C, Adams ET, Olson GR, Witt W, et al.. Utilizing Whole Slide Images for Pathology Peer Review and Working Groups. Toxicologic pathology (Print). 2015. DOI: https://doi.org/10.1177/0192623315605933

[34] Owembabazi E, Usman I, Makena W. Segmentation and quantification of testicular histology images using machine learning bioimage analysis tools; Ilastik and Fiji software. MethodsX. 2025. DOI: https://doi.org/10.1016/j.mex.2025.103503

[35] Paul AS, Rodrigues L, Kumar SKC, Manthey D, Border SP, Pardinhas C, et al.. ComPRePS: An Automated Cloud-based Image Analysis tool to democratize AI in Digital Pathology. bioRxiv. 2024. DOI: https://doi.org/10.1101/2024.03.21.586102

[36] Mercan E, Mehta S, Bartlett J, Weaver D, Elmore J, Shapiro L. Automated Diagnosis of Breast Cancer and Pre-invasive Lesions on Digital Whole Slide Images. International Conference on Pattern Recognition Applications and Methods. 2018. DOI: https://doi.org/10.5220/0006550600600068

[37] Neto PC, Oliveira SP, Montezuma D, Fraga J, Monteiro A, Ribeiro L, et al.. iMIL4PATH: A Semi-Supervised Interpretable Approach for Colorectal Whole-Slide Images. Cancers. 2022. DOI: https://doi.org/10.3390/cancers14102489

[38] Zhou H, Cui L. Integrating AI-Powered Digital Pathology With Case-Based Teaching: A Novel Paradigm for Renal Education in Medical School. The Clinical Teacher. 2026. DOI: https://doi.org/10.1111/tct.70421

[39] Qiu J, Liu Y. Multi-Directional Context Modeling With HCSMIL: Enhancing Cancer Prediction and Subtype Classification From Whole Slide Images. International journal of imaging systems and technology (Print). 2026. DOI: https://doi.org/10.1002/ima.70287

[40] Mukhopadhyay S, Dasgupta T, Walsh E, Millican-Slater R, Hanby A, Stephenson J, et al.. Abstract P3-05-48: Prediction of disease recurrence in low risk Oncotype Dx breast cancers from digital H&E-stained whole slide images of pre-treatment resections alone. Cancer Research. 2023. DOI: https://doi.org/10.1158/1538-7445.sabcs22-p3-05-48

[41] Jiang P, Liu J, Wang L, Feng J, Cao D, Pang B. Classifying Cervical Histopathological Whole Slide Images via Deep Multi-Instance Transfer Learning. IEEE International Conference on Bioinformatics and Biomedicine. 2022. DOI: https://doi.org/10.1109/BIBM55620.2022.9995014

[42] Lu M, Wu JC, Lu HHS, Eslam M, Yu M. Artificial Intelligence Applications in the Diagnosis, Treatment, and Prognosis of Hepatocellular Carcinoma. Gut and Liver. 2025. DOI: https://doi.org/10.5009/gnl250268

[43] Cho WC, Gill P, Aung P, Gu J, Nagarajan P, Ivan D, et al.. The utility of digital pathology in improving the diagnostic skills of pathology trainees in commonly encountered pigmented cutaneous lesions during the COVID-19 pandemic: A single academic institution experience. Annals of Diagnostic Pathology. 2021. DOI: https://doi.org/10.1016/j.anndiagpath.2021.151807

[44] Harazono Y, Fukawa Y, Iwasaki T, Hama Y, Nishii N, Tanaka Y, et al.. Multicenter clinicopathological study of odontogenic myxoma spectrum lesions using quantitative pathology. Scientific Reports. 2026. DOI: https://doi.org/10.1038/s41598-026-42019-8

[45] Basety S, Gudepu R, Velidandi A. Artificial Intelligence in Lung Cancer: A Narrative Review of Recent Advances in Diagnosis, Biomarker Discovery, and Drug Development. Pharmaceutics. 2026. DOI: https://doi.org/10.3390/pharmaceutics18020201

[46] Zhao Y, Marode TP, Bhangdiya VK, Nemane SG, Tulaskar DP, Sarad VM, et al.. Artificial Intelligence Meets Nail Diagnostics: Emerging Image-Based Sensing Platforms for Non-Invasive Disease Detection. Bioengineering. 2026. DOI: https://doi.org/10.3390/bioengineering13010075

[47] Wang J, Xu J, Zhou Y, Luo X, Wang H, Wang T, et al.. MultiXpert: Dual-stream synergistic enhancement with cross-modal alignment for zero-shot chest x-ray diagnosis . Information Processing & Management. 2026. DOI: https://doi.org/10.1016/j.ipm.2025.104468