Whole Slide Image Processing: From Acquisition to Analysis

Whole slide image processing is the end-to-end workflow that converts glass histopathology slides into digital files suitable for computational analysis. This workflow spans tissue scanning, file storage, preprocessing, and algorithmic interpretation, with each stage introducing decisions that affect downstream results. For students, researchers, analysts, and life-science professionals, understanding this pipeline is essential because the choices made during acquisition and preparation determine whether subsequent analyses are reproducible, interoperable, and clinically meaningful. This article describes the complete workflow, the practical decisions at each stage, and the quality controls needed to produce reliable results.

The Scope of Whole Slide Image Processing

Whole slide imaging transforms physical tissue sections on glass slides into high-resolution digital representations that can be viewed, shared, and analyzed on computer systems. The resulting whole slide images (WSIs) are exceptionally large, often reaching several gigabytes when uncompressed, which creates unique challenges for storage, processing, and analysis [5]. Unlike conventional digital images that can be loaded entirely into memory, WSIs require specialized handling because their size exceeds the capacity of typical computational resources.

The processing pipeline for WSIs involves multiple interconnected stages. Scanning captures the tissue at high magnification, producing a pyramid-structured file with multiple resolution levels. Storage systems must accommodate the large file sizes while maintaining fast access for viewing and analysis. Preprocessing prepares the images for computational methods through steps such as artifact detection, color normalization, and tiling. Analysis applies algorithms to extract quantitative information, ranging from simple tissue detection to complex deep learning models for disease classification [5].

Each stage presents distinct challenges. The incorporation of deep learning algorithms into routine workflow requires many more steps than simply training and testing a model [5]. Images must be prepared for modeling, and the resultant predictions need further processing for interpretation. Examples include artifact detection, color normalization, image subsampling or tiling, and removal of errant predictions [5]. Understanding these steps is critical for anyone working with WSI data.

At a Glance: Whole Slide Image Processing Stages

The following table summarizes the primary stages of the whole slide image processing workflow, the key decisions at each stage, and the practical considerations that affect outcomes.

Scanning and Acquisition

The scanning stage converts physical glass slides into digital files. Whole-slide scanners vary in technical features that affect magnification, throughput, image quality, and resulting file formats and sizes [14]. The choice of scanner influences every subsequent stage of the pipeline, making acquisition decisions critically important.

Scanner Selection and Tissue Detection

Whole-slide imaging systems differ in their tissue detection algorithms, which can alter the scanned area, scan time, and file size [14]. In high-volume laboratories, these differences translate into tangible workflow and cost implications. Scan area, determined by scanner-specific tissue detection algorithms, directly influences scan time, which in turn affects workflow and staff planning as well as file size, a major expense in storage and data management [14].

A comparative analysis of tissue detection patterns across four commercial whole-slide scanner models evaluated their effects on these metrics using classical computer vision to establish perimeter-based benchmarks [14]. The study scanned 260 routine diagnostic glass slides, balanced by hematoxylin and eosin-stained and immunohistochemistry slide types and biopsy or resection tissue types, on 8 whole-slide scanners representing 4 different commercial manufacturers [14]. The results demonstrated that scanner area, scan time, and file size extracted from WSI metadata varied significantly between scanner models, with linear regression quantifying the relationship between area and downstream metrics [14].

For practitioners, this means that scanner selection should consider also image quality but also operational efficiency. A scanner that produces larger scan areas than necessary will increase scan time and file size, consuming more storage and slowing throughput. Understanding the tissue detection behavior of a scanner before committing to a high-volume workflow is essential.

Scanning Quality and Verification

Verification of scanner performance is a critical step before deploying digital pathology for routine diagnostic use. A large multicenter study involving seven hospitals across Wales validated the Aperio AT2 scanner for reporting on a digital pathology platform [18]. The study involved 22 pathologists and scanned 7,491 slides from 3,001 cases, with a senior pathology fellow comparing digital reports with authorized reports on glass slides [18]. A panel of expert pathologists reviewed discrepant cases under a multiheader microscope to establish ground truth [18].

The results showed that 2,745 out of 3,001 cases (91%) demonstrated complete concordance between digital and glass slide reports [18]. Two hundred and fifty-six cases showed discrepancies in diagnosis, of which 170 (5.6%) were deemed of no clinical significance by the review panel [18]. There were 86 (2.9%) clinically significant discrepancies, with ground truth lying with digital pathology in 28 cases and with glass slides in 58 cases [18]. The sensitivity of digital pathology was 98.07% compared to 99.07% for glass slides [18].

This study demonstrates that scanner verification is a formal process requiring structured comparison against established methods. For research settings, similar verification should be performed before relying on scanned images for quantitative analysis.

Storage and Data Management

Whole slide images present significant data storage challenges. Digital pathology WSIs are large images, approximately 30 GB per slide when uncompressed, with high resolution at 0.25 microns per pixel [21]. Hospitals and research institutions adopting digital pathology must address this storage burden through compression and efficient data management strategies.

File Formats and Compression

Lossy compression has been adopted by scanner manufacturers to address the storage challenge [21]. A comparison of lossy JPEG compression for WSIs investigated the Vector Quantised Variational Autoencoder 2 variant (VQVAE2) as a possible alternative to reduce file size while encoding useful features in the compressed representation [21]. Three VQVAE2 models were trained on a Camelyon 2016 subset to compression ratios of 19.2:1, 9.6:1, and 4.8:1 and tested on multiple datasets [21]. Both JPEG and JPEG 2000 compression outperformed the VQVAE2 implementation within Peak Signal to Noise Ratio and Structural Similarity metrics [21]. The trained VQVAE2 models could visually reproduce WSI tissue structure but used colors from the original training data within the reconstructions on other datasets [21].

For practical purposes, this means that established compression methods such as JPEG and JPEG 2000 remain reliable choices for WSI storage. Newer compression approaches may offer visual quality but can introduce color shifts that affect downstream analysis.

High-Performance Data Management

When dealing with giga-pixel digital pathology in whole-slide imaging, a notable proportion of data records holds relevance during each analysis operation [11]. When deploying an image analysis algorithm on WSIs, the computational bottleneck often lies in the input-output (I/O) system [11]. This is particularly notable as patch-level processing introduces a considerable I/O load onto the computer system [11].

The Adaptable IO System version 2 (ADIOS2) has been implemented as a digital pathology-centric pipeline to facilitate streamlined data management across WSIs [11]. Performance evaluation encompassed two key scenarios: a pure CPU-based image analysis scenario and a GPU-based deep learning framework scenario [11]. Under the CPU scenario, ADIOS2 showcased an impressive two-fold speed-up compared to the brute-force approach [11]. In the GPU scenario, its performance stood on par with the cutting-edge GPU I/O acceleration framework, NVIDIA Magnum IO GPU Direct Storage [11].

This finding highlights that data management is also a storage concern but a computational performance factor. The I/O system can become the bottleneck in WSI analysis, and optimizing data access patterns can substantially reduce processing time.

Preprocessing for Analysis

Preprocessing prepares whole slide images for computational analysis. Deep learning methods often require substantial pre- and post-processing to improve interpretation and prediction [5]. Similar to any data processing pipeline, images must be prepared for modeling, and the resultant predictions need further processing for interpretation [5]. Examples include artifact detection, color normalization, image subsampling or tiling, and removal of errant predictions [5].

Artifact Detection

Artifacts introduced during slide preparation and whole-slide imaging can compromise diagnostic accuracy [13]. Quality control remains a major challenge in digital histopathology, and automated QC has emerged as a potential solution to the scalability and variability limitations of manual review in digital pathology workflows [13].

A clinical implementation of a commercially available AI-based automated QC platform was evaluated within a large digital pathology laboratory [13]. Feasibility testing included 60 histopathology slides scanned across 3 WSI platforms, followed by a retrospective operational analysis on 94,995 WSIs generated over 1 month across 10 subspecialty services and multiple stain types [13]. Artifact detection demonstrated high reproducibility across scanner platforms, with tissue fold detection showing greater than 97% concordance and air bubble detection exceeding 98% concordance [13]. Scanning-related artifacts, including missing tissue and blurred regions, varied by scanner model but were consistently identified [13]. Across the clinical deployment, the mean analysis time was 17 seconds per image, with no analysis failures observed during load testing [13]. The mean total artifact burden per image was 1.91%, with most images meeting predefined QC thresholds [13].

For research workflows, automated artifact detection should be considered a standard preprocessing step. Manual review of every WSI is impractical at scale, and automated systems can consistently identify common artifacts across different scanner platforms.

Color Normalization and Stain Standardization

Pre-analytical factors like how the tissue is prepared, staining, and scanning artifacts affect AI applications and contain possible post-analytical barriers such as the range of colors used and color standardization [22]. Stain variation between laboratories, batches, and individual slides can significantly affect algorithmic performance.

A fast and scalable pipeline for stain normalization of whole-slide images in histopathology has been described in the literature [23]. Stain normalization aims to reduce color variation while preserving tissue structure, making images more comparable across different sources. For researchers developing algorithms, stain normalization should be applied before model training and inference to improve generalization across datasets.

Tiling and Patch Extraction

Once processed, predictions are complicated by image file size, typically several gigabytes when unpacked [5]. This forces images to be tiled, meaning that a series of subsamples from the whole-slide image are used in modeling [5]. Tiling is a fundamental preprocessing step that enables computational analysis of WSIs.

A general and practical pipeline for glomerulus segmentation demonstrated that stitching on overlapping patches increases detection coverage, especially when glomeruli are located near patch image borders [6]. The approach was evaluated across two large and diverse datasets with over 30,000 glomerulus annotations, and models using the pipeline outperformed the previous state-of-the-art method [6].

The choice of patch size and overlap strategy affects both detection performance and computational efficiency. Overlapping patches reduce the risk of missing structures at patch boundaries but increase the number of patches to process. For applications where detection of small structures is critical, overlapping patches should be considered.

Analysis Methods and Algorithms

The analysis stage applies computational methods to extract quantitative information from preprocessed WSIs. The scope of AI applications in pathology includes automated image capture, assessment and analysis, risk stratification, and prognostic prediction [22]. The integration of WSI and AI introduces significant challenges, including data quality, high computational demands, the ability to generalize across different settings, and a range of ethical considerations [22].

Deep Learning for Segmentation and Classification

Deep learning has pushed the scope of digital pathology beyond simple digitization and telemedicine [5]. The incorporation of these algorithms in routine workflow is on the horizon and may be a disruptive technology, reducing processing time and increasing detection of anomalies [5].

An automated machine learning-based pipeline for obtaining 11-component differential cell counts on whole-slide bone marrow aspirate smears demonstrated the feasibility of fully automated analysis [12]. The pipeline used a sequential process of identifying optimal bone marrow aspirate regions with high proportions of marrow nucleated cells, detecting individual cells within these optimal areas, and classifying these cells into 1 of 11 differential cell count components [12]. Convolutional neural network models were trained on 396,048 bone marrow aspirate region images, 28,914 cell boundary images, and 1,510,976 cell class images from manual annotations [12]. The resulting automated pipeline produced differential cell counts that demonstrated high statistical and diagnostic concordance with manual counts among a heterogeneous group of testing slides with varying pathologies and cellularities [12]. Additionally, automated analysis reduced the intraslide variance in differential cell counts by analyzing the whole slide and marrow nucleated cells within all optimal regions [12].

This example illustrates the potential of deep learning pipelines to replace manual analysis while improving consistency. The pipeline outputs of region classification, cell detection, and cell classification can be visualized using whole-slide image analysis software [12].

Foundation Models and Zero-Shot Segmentation

The Segment Anything Model (SAM) was released as a foundation model for image segmentation, trained by over 1 billion masks on 11 million licensed and privacy-respecting images [20]. The model supports zero-shot image segmentation with various segmentation prompts such as points, boxes, and masks [20]. An evaluation of SAM on representative segmentation tasks on WSI, including tumor segmentation, non-tumor tissue segmentation, and cell nuclei segmentation, found that the zero-shot SAM model achieves remarkable segmentation performance for large connected objects [20]. However, it does not consistently achieve satisfying performance for dense instance object segmentation, even with 20 prompts on each image [20].

Identified limitations for digital pathology include image resolution, multiple scales, prompt selection, and model fine-tuning [20]. Few-shot fine-tuning with images from downstream pathological segmentation tasks might help the model achieve better performance in dense object segmentation [20].

For researchers considering foundation models, the key practical implication is that zero-shot performance may be adequate for large structures but insufficient for dense cellular analysis. Fine-tuning on domain-specific data is likely necessary for challenging segmentation tasks.

Quantitative Analysis Pipelines

Multimodal and multiscale imaging is crucial for quantitative high-content spatial profiling [9]. An integrated image processing pipeline for comprehensive tissue analysis combines quantitative phase microscopy for tissue architecture mapping, hyper-plex fluorescence imaging for immune microenvironment profiling, and whole-slide histopathology [9]. This approach enables detailed morphological mapping of tissue architecture and cell morphology while simultaneously linking them to the functional states of individual cells across the entire slide [9]. By analyzing tissue biopsies from patients with ulcerative colitis, the pipeline demonstrated potential for quantitative spatial analysis of molecular markers related to mucosal healing [9]. The pipeline is open-source and compatible with conventional microscopy systems [9].

A quantitative image analysis-based pipeline called CRCS was proposed to explore the relationship between the expression level of adreno-cortico-tropic-hormone in normal cell tissues adjacent to tumor cells and the postoperative prognosis of patients with Cushing's disease [10]. CRCS mainly consists of image-level clustering, cluster-level multi-modal image registration, patch-level image classification, and pixel-level image segmentation on whole slide imaging [10]. On both image registration and classification tasks, CRCS achieved state-of-the-art performance compared to recently published methods on a collected benchmark dataset [10]. CRCS achieved an accuracy of 0.83 for postoperative prognosis of 12 cases [10].

These examples demonstrate that WSI analysis pipelines can be tailored to specific clinical questions, integrating multiple imaging modalities and analysis steps to produce clinically relevant outputs.

Workflow Implementation and Assessment

Implementing a whole slide image processing workflow requires structured planning and validation. The following steps provide a practical framework for establishing a pipeline in a research or clinical setting.

Step 1: Define the Analysis Question and Required Outputs

Before selecting scanners, storage systems, or algorithms, define the specific question the analysis must answer. Determine whether the output is a classification label, a quantitative measurement, or a spatial map. This decision determines the required image resolution, the need for specialized staining, and the choice of analysis algorithms.

Step 2: Select and Verify Scanning Infrastructure

Choose a scanner appropriate for the tissue types and analysis requirements. Verify scanner performance using a representative set of slides before committing to large-scale scanning. Document the tissue detection behavior of the scanner, as this affects scan area, scan time, and file size [14]. For clinical applications, formal verification against established methods is necessary [18].

Step 3: Establish Storage and Data Management Systems

Plan storage capacity based on expected slide volumes and file sizes. Consider compression methods that balance file size against image quality for downstream analysis [21]. Evaluate data management systems that optimize I/O performance, as patch-level processing can create a computational bottleneck [11].

Step 4: Implement Preprocessing Controls

Develop a preprocessing pipeline that includes artifact detection, color normalization, and tiling. Automated QC systems can consistently identify common artifacts across scanner platforms [13]. Stain normalization should be applied to reduce color variation between slides and laboratories [23].

Step 5: Select and Validate Analysis Algorithms

Choose algorithms appropriate for the analysis question. Validate algorithm performance on a representative dataset with known ground truth. For deep learning approaches, ensure that training data adequately represents the range of conditions expected in deployment.

Step 6: Document and Report Results

Maintain detailed records of all pipeline parameters, including scanner settings, preprocessing steps, and algorithm versions. This documentation is essential for reproducibility and for interpreting results in the context of the specific workflow used.

Records and Measurements

Maintaining accurate records is essential for reproducible whole slide image processing. The following measurements should be documented for each slide processed through the pipeline.

Acquisition Records

Record the scanner model, scanning magnification, tissue detection algorithm behavior, scan area, scan time, and file size for each slide [14]. These parameters affect downstream processing and should be tracked to identify systematic variations between scanner models or batches.

Quality Control Records

Document artifact detection results, including the types and locations of artifacts identified. Record the artifact burden per image and whether the image met predefined quality thresholds [13]. Track processing latency and any analysis failures during load testing [13].

Preprocessing Records

Document the preprocessing steps applied to each image, including color normalization parameters, tiling strategy, and patch overlap settings. Record the software versions and configuration files used for each preprocessing step.

Analysis Records

Record the algorithm versions, model weights, and inference parameters used for analysis. Document the computational resources required, including processing time and I/O performance [11]. For deep learning models, record the training data composition and any fine-tuning steps applied.

Common Failure Patterns

Understanding common failure patterns in whole slide image processing helps practitioners anticipate and prevent problems.

Scanner-Related Failures

Scanner-specific tissue detection algorithms can produce scan areas that are larger or smaller than the minimal tissue perimeter [14]. This variation affects scan time and file size, creating inefficiencies in high-volume workflows. Scanning-related artifacts, including missing tissue and blurred regions, vary by scanner model [13]. Regular monitoring of artifact prevalence by scanner model can identify systematic issues.

Compression-Related Failures

Lossy compression can introduce color shifts that affect downstream analysis. The VQVAE2 compression approach used colors from the original training data within reconstructions on other datasets, demonstrating that compression methods can introduce dataset-specific biases [21]. Validate compression approaches on the specific tissue types and stains used in the workflow.

Model Generalization Failures

Deep learning models may not generalize across different settings, staining protocols, or scanner platforms [22]. Bias found in training datasets can blur ethical and legal boundaries alongside regulatory uncertainty [22]. Validate models on diverse datasets that represent the range of conditions expected in deployment.

I/O Bottlenecks

Patch-level processing introduces considerable I/O load onto computer systems [11]. Without optimized data management, the I/O system becomes the computational bottleneck, slowing analysis and reducing throughput. Evaluate data management systems that parallelize patch-level processing [11].

Limitations and Interpretation

Whole slide image processing has inherent limitations that affect the interpretation of results.

Resolution and Scale Limitations

Foundation models for segmentation may not consistently achieve satisfying performance for dense instance object segmentation, even with multiple prompts [20]. Identified limitations include image resolution and multiple scales [20]. For applications requiring detection of individual cells or small structures, specialized models may be necessary.

Concordance Limitations

While automated analysis can achieve high concordance with manual assessment, agreement may be weaker for certain cell-level outputs [17]. A deep learning-based framework for quantitative assessment of endoscopic biopsies in Crohn's disease showed good concordance with pathologist assessment for crypt abscess and cryptitis ratios, but agreement was weaker for certain cell-level outputs, particularly submucosal plasma cell density [17]. Grade-stratified calibration indicated a tendency toward AI over-detection in submucosal regions with negative or low manual scores, underscoring the need for cautious interpretation of this metric [17].

De-identification Requirements

De-identification of medical images intended for research is a core requirement for data sharing initiatives, particularly as the demand for data for AI applications grows [8]. The National Cancer Institute convened a virtual workshop to summarize the state of the art in de-identification technology and processes, covering pathology whole slide image de-identification, de-facing, the role of AI in image de-identification, and the NCI Medical Image De-Identification Initiative datasets and pipeline [8]. Researchers sharing WSI data must implement appropriate de-identification procedures.

Data Sharing and FAIR Principles

Data sharing initiatives require attention to the FAIR Guiding Principles, which provide a framework for making data findable, accessible, interoperable, and reusable [4]. The Genomic Data Sharing Policy from the National Institutes of Health provides requirements for data sharing in genomic research [3]. While these policies focus on genomic data, the principles apply to WSI data sharing in research contexts. Resources such as EMBL-EBI Training and NCBI Data Resources provide guidance on data management and sharing practices [1][2].

Safety and Regulatory Context

Whole slide image processing in clinical settings operates within a regulatory framework that requires validation and quality assurance.

Clinical Validation

Verification and validation of digital pathology for primary histopathological diagnosis requires formal processes. The All Wales experience demonstrated that scanner verification and pathologist validation are essential before deploying digital pathology for routine diagnostic services [18]. The study concluded that the scanner produced adequate quality images for routine histopathologic diagnosis and that pathologists were able to diagnose with good concordance as with glass slides [18].

Quality Control Standards

Automated QC systems must demonstrate reproducibility across scanner platforms and maintain stable performance under sustained high-throughput conditions [13]. For clinical deployment, the system must meet predefined quality thresholds and process images within acceptable latency limits [13].

Ethical Considerations

The integration of AI in digital pathology introduces ethical considerations, including data quality, high computational demands, the ability to generalize across different settings, and regulatory uncertainty [22]. Researchers and clinicians must consider these factors when implementing WSI analysis pipelines.

Professional Escalation Criteria

Knowing when to escalate issues to specialized personnel is critical for maintaining workflow quality.

Escalate When Artifact Burden Exceeds Quality Thresholds

If automated QC identifies artifact burdens that exceed predefined thresholds, escalate to laboratory personnel for review. Artifacts such as tissue folds, air bubbles, missing tissue, and blurred regions can compromise diagnostic accuracy [13].

Escalate When Model Performance Degrades

If model performance on validation datasets falls below acceptable levels, escalate to computational pathology specialists. Performance degradation may indicate issues with staining, scanning, or model generalization that require specialized expertise.

Escalate When Concordance Is Poor

If automated analysis shows poor concordance with manual assessment for specific metrics, escalate to pathologists for review. Certain cell-level outputs may require cautious interpretation due to tendencies toward over-detection [17].

Escalate When Data Sharing Requirements Are Unclear

If data sharing requirements are unclear, escalate to institutional data governance personnel. De-identification of medical images is a core requirement for data sharing initiatives [8], and compliance with applicable policies is essential [3].

Frequently Asked Questions

What is a whole slide image sample?

A whole slide image sample is a digital representation of an entire glass histopathology slide, captured at high resolution by a specialized scanner. The resulting file contains the complete tissue section at multiple magnification levels, allowing viewers to navigate from low-power overview to high-power detail. Whole slide images are typically several gigabytes when uncompressed, requiring specialized storage and processing approaches [5][21].

What is whole slide analysis?

Whole slide analysis refers to the computational examination of entire whole slide images instead of selected regions. This approach can reduce intraslide variance by analyzing all optimal regions of the slide instead of a small subset [12]. Whole slide analysis encompasses tasks such as tissue segmentation, cell detection and classification, and quantitative measurement of histological features [12][17].

What is lazyslide accessible and interoperable whole slide image analysis?

Lazyslide refers to an approach to whole slide image analysis that emphasizes accessibility and interoperability. The term describes software and workflows designed to work with standard WSI formats and to be usable across different computing environments. Interoperability in computational pathology requires alignment with standardized hierarchical information models, such as those defined by DICOM, to enable transparent dataset characterization and improve reproducibility [16].

How long does whole slide image processing take?

Processing time varies by pipeline stage and computational infrastructure. Automated quality control systems have demonstrated mean analysis times of 17 seconds per image in clinical deployment [13]. Analysis time depends on the complexity of the algorithms, the size of the images, and the efficiency of the data management system. Optimized I/O systems can provide substantial speed-ups compared to brute-force approaches [11].

What causes variation in whole slide image file sizes?

File size variation is primarily caused by differences in scan area, which is determined by scanner-specific tissue detection algorithms [14]. Scanners that detect larger tissue areas produce larger files. Compression methods also affect file size, with lossy compression reducing storage requirements at the cost of some image quality [21].

How should I choose between different whole slide scanners?

Scanner selection should consider tissue detection behavior, image quality, throughput, and file size implications [14]. Verify scanner performance on representative slides before committing to high-volume workflows. For clinical applications, formal verification against established methods is required [18].

What preprocessing steps are essential for deep learning analysis?

Essential preprocessing steps include artifact detection, color normalization, and tiling [5]. Artifact detection identifies tissue folds, air bubbles, missing tissue, and blurred regions that can compromise analysis [13]. Color normalization reduces stain variation between slides [23]. Tiling divides the large WSI into manageable patches for modeling [5].

How do I validate a whole slide image analysis pipeline?

Validation requires comparison against established methods using representative datasets. For clinical applications, this involves structured comparison of digital reports against glass slide reports with expert review of discrepancies [18]. For research applications, validate model performance on diverse datasets that represent the range of conditions expected in deployment [22].

Related Bioinformatics Guides

• Cryo-EM Image Processing and 3D Reconstruction
• Pan-Cancer Analysis of Whole Genomes (PCAWG)
• Gene Ontology (GO) and Enrichment Analysis
• The KEGG Database and Pathway Analysis
• Docker and Containerization in Reproducible Research

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Developing image analysis pipelines of whole-slide images: Pre- and post-processing.. Journal of clinical and translational science, 2020.
• A General Pipeline for Glomerulus Whole-Slide Image Segmentation.. Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE Engineering in Medicine and Biology Society. Annual International Conference, 2025.
• Open-Source Whole Slide Image Preparation and Viewing Pipeline.. Archives of pathology & laboratory medicine, 2018.
• Summary of the National Cancer Institute 2023 Virtual Workshop on Medical Image De-identification-Part 2: Pathology Whole Slide Image De-identification, De-facing, the Role of AI in Image De-identification, and the NCI MIDI Datasets and Pipeline.. Journal of imaging informatics in medicine, 2025.
• Multimodal whole slide image processing pipeline for quantitative mapping of tissue architecture and tissue microenvironment.. Npj imaging, 2025.
• CRCS: An automatic image processing pipeline for hormone level analysis of Cushing's disease.. Methods (San Diego, Calif.), 2024.
• High-performance Data Management for Whole Slide Image Analysis in Digital Pathology.. Proceedings of SPIE--the International Society for Optical Engineering, 2024.
• An Automated Pipeline for Differential Cell Counts on Whole-Slide Bone Marrow Aspirate Smears.. Modern pathology : an official journal of the United States and Canadian Academy of Pathology, Inc, 2023.
• Deployment of AI-driven automated quality control of whole-slide images in a large tertiary cancer center.. 2026.
• Comparative analysis of whole-slide scanner tissue detection algorithms: Implications for scan area, scan time, and file size in high-volume digital pathology workflows.. 2026.
• Computer-aided image analysis for PD-L1 assessment in canine urothelial carcinoma.. 2026.
• A straightforward framework to harmonize computational pathology.. 2026.
• Deep learning-based quantitative histopathology of endoscopic biopsies in Crohn's disease: a retrospective cross-sectional validation study.. 2026.
• Verification and Validation of Digital Pathology (Whole Slide Imaging) for Primary Histopathological Diagnosis: All Wales Experience. Journal of Pathology Informatics, 2021.
• Digital pathology & whole slide imaging: A viewpoint. IP Archives of Cytology and Histopathology Research, 2025.
• Segment Anything Model (SAM) for Digital Pathology: Assess Zero-shot Segmentation on Whole Slide Imaging. IS&T International Symposium on Electronic Imaging, 2023.
• Digital pathology whole slide image compression with vector quantized variational autoencoders. Medical Imaging, 2023.
• The Role of Whole Slide Imaging in AI-Based Digital Pathology: Current Challenges and Future Directions-An Updated Literature Review. Journal of Molecular Pathology, 2026.
• A fast and scalable pipeline for stain normalization of whole-slide images in histopathology. Lecture Notes in Computer Science Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics, 2019.
• Accelerating Artificial Intelligence-based Whole Slide Image Analysis with an Optimized Preprocessing Pipeline. Informatik Aktuell, 2024.
• Ink removal from histopathology whole slide images by combining classification, detection and image generation models. Proceedings International Symposium on Biomedical Imaging, 2019.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.