Whole Slide Image Analysis: A Practical Workflow for Pathologists

Whole slide image analysis converts glass histopathology slides into digital files that can be viewed, annotated, and processed with computational tools. This workflow spans slide scanning, image quality control, tissue detection, preprocessing, model application, and result interpretation. For students, researchers, analysts, and life-science professionals, the practical challenge is building a reproducible pipeline that handles gigapixel images without losing diagnostic or research value. This article provides a step-by-step workflow grounded in published methods and validated implementations, with attention to data management, quality checks, and the limits of computational interpretation.

At a Glance

The table below summarizes the core stages of a whole slide image analysis workflow, the primary decisions at each stage, and the records you should keep for reproducibility.

The Scope of Whole Slide Image Analysis

Whole slide imaging has moved pathology from microscope-based review to a computer vision workflow. Digitized slides can be annotated and analyzed with algorithms that detect structures, classify tissue patterns, and quantify biomarkers. The practical value depends on the entire pipeline, also the model. Deep learning has expanded digital pathology beyond simple digitization and telemedicine, and incorporating these algorithms into routine laboratory workflow requires substantial pre- and post-processing to improve interpretation and prediction. 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. Once processed, predictions are complicated by image file size, typically several gigabytes when unpacked, which forces images to be tiled into a series of subsamples for modeling. This review of methods for biopsy slide analysis discusses the unique issues that come with analyzing very large images. See the EMBL-EBI Training resources for foundational instruction on bioimage analysis and data handling.

The workflow described here applies to research settings, translational studies, and clinical validation contexts. The same principles govern a research project quantifying pancreatic islet cell mass and a diagnostic validation study comparing digital reporting with glass slide review. What changes is the rigor required at each step and the documentation expected by reviewers or regulators.

Slide Scanning and Image Acquisition

Scanner Selection and Tissue Detection

Whole-slide scanners vary in technical features that affect magnification, throughput, image quality, and resulting file formats and sizes. A comparative analysis of four commercial scanner models found that tissue detection algorithms differ across systems, and these differences alter the scanned area, scan time, and file size. 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 affects workflow and staff planning, as well as file size, a major expense in storage and data management. The study used 260 routine diagnostic glass slides balanced by hematoxylin and eosin-stained and immunohistochemistry slide types and biopsy or resection tissue types, scanned on eight whole-slide scanners representing four manufacturers. A classical computer vision pipeline delineated the minimal tissue perimeter on each slide as the reference area, and scanner area, scan time, and file size were extracted from the whole slide image metadata. See the comparative analysis of whole-slide scanner tissue detection algorithms for the full methodology.

For your own workflow, record the scanner model, the tissue detection setting, and the resulting scan area for each batch. If you observe that one scanner consistently produces larger files for the same tissue type, investigate whether the tissue detection algorithm is including pen marks, bubbles, or glass debris in the scan region. Adjust the detection sensitivity or use manual scan region selection when the automated detection fails.

Scan Quality Verification

A large multicenter study across seven hospitals in Wales validated the Leica Aperio AT2 scanner for routine diagnostic histopathological services. The study involved 22 pathologists, 7491 slides from 3001 cases, and compared digital pathology reports with authorized glass slide reports. Complete concordance between digital and glass slide reports occurred in 91% of cases. Of 256 discrepant cases, 170 were deemed of no clinical significance, and 86 were clinically significant. Ground truth lay with digital pathology in 28 of the clinically significant discrepancies and with glass slide in 58 cases. Sensitivity of digital pathology was 98.07% and for glass slide was 99.07%. The authors concluded that the scanner produced adequate quality images for routine histopathologic diagnosis and that pathologists could diagnose with good concordance using digital pathology. See the All Wales digital pathology verification study for details.

The practical lesson is that scanner verification is a prerequisite for trusting downstream analysis. Before you run any computational model on scanned slides, confirm that the scanner produces images adequate for the intended diagnostic or research purpose. This means checking focus quality, color fidelity, and resolution on a representative sample of slides from each tissue type and stain you plan to analyze.

Quality Control and Artifact Management

Artifact Detection Before Model Application

Reliable whole slide image analysis depends on diagnostic models, but it also depends on image quality control that can be assessed before those models are applied. High-sensitivity artifact masks may remove relevant tissue, especially in densely nucleated regions where benign histology resembles dust, folds, or preparation debris. A conservative preprocessing pipeline called GuardQC was developed for artifact control and tissue preservation in computational pathology. The method uses a coarse-to-fine strategy: a high-recall proposal network identifies suspicious regions, and a dual-stream ResNet-18 verification model combines local texture with tissue context after dynamic multi-resolution alignment. Iterative hard negative mining improves specificity under noisy quality control labels. Localized Otsu re-segmentation contracts confirmed artifact regions, and deterministic PDE-based inpainting neutralizes artifacts without synthesizing new tissue content. On an expert-curated subset, GuardQC achieved 98.99% accuracy and 100% specificity with zero false positives. In the evaluated scanner-stratified whole slide image cohort, the masked tissue percentage was 0.32%, compared with 0.94% for a deep learning baseline. See the GuardQC artifact quality control study for the full evaluation.

The key decision for your workflow is whether to mask artifacts or inpaint them. Masking removes artifact regions from analysis, which preserves tissue but may exclude relevant areas. Inpainting neutralizes artifacts without removing tissue, but you must verify that the inpainting does not synthesize new tissue content that could mislead your model. For research workflows, masking is the more conservative choice. For clinical workflows where the entire slide must be assessed, inpainting may be necessary, but you should document the inpainting method and validate that it does not alter diagnostic features.

Tissue Preservation Versus Aggressive Masking

The tension between artifact removal and tissue preservation is a recurring theme in whole slide image analysis. Aggressive artifact detection can remove densely nucleated regions where benign histology resembles dust, folds, or preparation debris. This is particularly problematic for tasks like tumor detection or metastasis classification, where the regions most likely to contain diagnostic information are also the regions most likely to trigger artifact detectors.

When you evaluate an artifact detection method, measure both sensitivity and the proportion of tissue masked. A method with high sensitivity but excessive tissue removal may reduce the usable area for analysis and bias your results. The GuardQC evaluation reported a masked tissue percentage of 0.32% compared with 0.94% for a deep learning baseline, demonstrating that conservative preprocessing can preserve tissue while maintaining artifact control. For your own pipeline, establish a threshold for acceptable masked tissue percentage based on your tissue type and analysis task. If the masked percentage exceeds your threshold, review the artifact detection settings and adjust the verification stage.

Preprocessing for Computational Analysis

Tiling and Subsampling

Whole slide images are typically several gigabytes when unpacked, which forces images to be tiled into a series of subsamples for modeling. Tiling is a necessary step because most deep learning models cannot process a full gigapixel image in a single pass. The tile size, overlap, and sampling strategy affect both model performance and computational cost.

An automated image preprocessing workflow for papillary thyroid carcinoma classification used whole slide images from The Cancer Genome Atlas repository and applied an automated tile extraction and preprocessing pipeline to ensure analysis-ready quality. The extracted image tiles were used to train a deep convolutional neural network, specifically Google's Inception v3, to classify tumors based on BRAFV600E or RAS mutations. The pipeline performed equally well as an expert-curated image classifier, achieving area under the curve values of 0.86 for validation and 0.865 for final testing. The classifier accurately predicted 90% of tumors in the validation set and 84.2% in the final testing set. See the papillary thyroid carcinoma WSI classification workflow for details.

When you design a tiling strategy, consider the following decisions:

• Tile size should match the receptive field of your model architecture. Common sizes range from 256x256 to 1024x1024 pixels.
• Overlap between tiles prevents information loss at tile boundaries. A 10% to 20% overlap is a reasonable starting point, but you should test whether overlap improves your model's performance.
• Sampling strategy matters for imbalanced datasets. If your slides contain mostly benign tissue with small tumor regions, random sampling may produce tiles that are overwhelmingly benign. Consider stratified sampling or oversampling of rare classes.
• Record the tile dimensions, overlap, and sampling strategy for every experiment so you can reproduce your results.

Color Normalization and Stain Variation

Staining techniques vary between laboratories, between batches, and even between slides in the same batch. Color normalization reduces this variation so that models trained on one set of slides can generalize to another. The preprocessing review from the Journal of Clinical and Translational Science lists color normalization as one of the essential pre-processing steps for whole slide image analysis, alongside artifact detection, image subsampling or tiling, and removal of errant predictions. See the pre- and post-processing review for WSI analysis pipelines for the full discussion.

The choice of color normalization method depends on your model and your data. Some methods match the color distribution of a reference slide, while others separate stain components and normalize each channel independently. For research workflows, document the normalization method and the reference slide used. For clinical workflows, verify that normalization does not obscure diagnostically relevant features.

Registration for Multiplexed and Multimodal Imaging

Some analyses require correlating measurements across several whole slide images of the same sample. This occurs in multiplexing, where the same sample is imaged multiple times through various systems, and in multimodal imaging, where the same system is used throughout rounds of sample manipulation such as multiple staining sessions. In both cases, slight deformations from one image to another are unavoidable, leading to imperfect superimposition and a loss of accuracy that makes it difficult to link measurements, particularly at the cellular level.

An open-source workflow using Fiji, QuPath, and Elastix facilitates the nonlinear registration of whole slide images to reach sub-cellular resolution. The set of whole slide images to register and analyze is defined as a QuPath project. Fiji is then used to open the QuPath project and perform the registrations. Each registration is automated using an elastix backend, or semi-automated using BigWarp to interactively correct the results of the automated registration. These transformations can be retrieved in QuPath to transfer regions of interest from one image to the corresponding registered images. See the open-source WSI registration workflow for the full protocol.

For your own multiplexed or multimodal analysis, verify registration accuracy at the cellular level before combining measurements across images. A registration error of a few pixels may be acceptable for tissue-level analysis but will corrupt cell-level measurements.

Model Selection and Training

Task-Specific Models Versus Foundation Models

Task-specific AI and machine learning models, such as convolutional neural networks, have achieved above-human performance in many image processing and computer vision tasks. However, the performance of task-specific models depends on the availability of many annotated training datasets, which presents a rate-limiting factor for AI and machine learning development in pathology. Task-specific models cannot benefit from multimodal data and lack generalization, often struggling to generalize to new datasets or unseen variations in image acquisition, staining techniques, or tissue types. See the review of generative AI and foundation models in digital pathology for the full discussion.

Foundation models are large AI models trained using sizable data, which are later adapted or fine-tuned to perform different tasks using a modest amount of task-specific annotated data. These models provide in-context learning, can self-correct mistakes, and promptly adjust to user feedback. The review proposes creating a pathology-specific generative AI based on multimodal foundation models and describes different use cases for its potentially transformative role in digital pathology.

For your workflow, the choice between a task-specific model and a foundation model depends on your annotated data availability and your generalization requirements. If you have a large annotated dataset for a specific task, a task-specific model may perform well. If you have limited annotations or need to generalize across staining variations and tissue types, a foundation model fine-tuned on your data may be more robust.

Handling Imbalanced Multiclassification

Sample imbalance remains an intractable challenge in whole slide image analysis for multiclassification tasks. A method called Cross-Patient Pseudo Bags Generation and Curriculum Contrastive Learning addresses this problem by learning fine-grained information through generating sub-bags with feature distributions similar to the original whole slide images. A pseudo-bag generation algorithm leverages the abundant and redundant information in whole slide images, allowing efficient training in unbalanced-sample multiclassification tasks. An affinity-based sample selection and curriculum contrastive learning strategy enhances the stability of model representation learning. The method demonstrated significant performance improvements on three datasets, including tumor classification and lymph node metastasis, achieving an average 4.39-point improvement in F1 score compared to the second-best method across the three tasks. See the cross-patient pseudo bags generation method for details.

When you face an imbalanced dataset, consider the following strategies:

• Generate pseudo-bags that preserve the feature distribution of the original whole slide images.
• Use curriculum learning to present easier samples first and gradually increase difficulty.
• Evaluate your model using F1 score or balanced accuracy instead of overall accuracy, which can be misleading with imbalanced classes.
• Document the class distribution in your training, validation, and test sets so reviewers can assess the generalizability of your results.

Model Validation and Performance Assessment

Benchmarking Against Pathologist Performance

A landmark study evaluated the performance of automated deep learning algorithms at detecting metastases in hematoxylin and eosin-stained tissue sections of lymph nodes of women with breast cancer and compared it with pathologists' diagnoses in a diagnostic setting. The CAMELYON16 challenge provided a training data set of whole-slide images from two centers in the Netherlands with and without nodal metastases verified by immunohistochemical staining. Algorithm performance was evaluated in an independent test set of 129 whole-slide images. The same test set of corresponding glass slides was also evaluated by a panel of 11 pathologists with time constraint from the Netherlands to ascertain likelihood of nodal metastases for each slide in a flexible 2-hour session, simulating routine pathology workflow, and by one pathologist without time constraint. The area under the receiver operating characteristic curve for the algorithms ranged from 0.556 to 0.994. See the CAMELYON16 deep learning algorithm assessment for the full results.

The wide range of algorithm performance in this study underscores the importance of rigorous validation. A model with an AUC of 0.556 is no better than chance, while a model with an AUC of 0.994 approaches perfect discrimination. The difference often lies in the quality of the training data, the preprocessing pipeline, and the validation methodology.

For your own model validation, follow these steps:

1. Split your data into training, validation, and test sets at the patient or case level, not the slide or tile level, to avoid data leakage.
2. Use an independent test set that was not used for any training decisions, including hyperparameter tuning.
3. Compare your model's performance against a relevant baseline, whether that is a pathologist's diagnosis, a published benchmark, or a simpler model.
4. Report confidence intervals for your performance metrics, beyond point estimates.
5. Document the inclusion and exclusion criteria for your test set.

Survival Analysis and Biomarker Development

Whole slide image analysis can extend beyond classification to predict clinical outcomes. A framework called WhARIO was developed for whole-slide-image-based survival analysis for patients treated with immunotherapy. The three-step framework combines contrastive learning and nonparametric clustering to distinguish tissue patterns within the slides, before exploiting the adjacencies of previously defined regions to derive features and train a proportional hazards model for survival analysis. The approach was tested on an in-house dataset of 193 patients from 5 medical centers and compared with the gold standard tumor proportion score biomarker. On fivefold cross-validation of the entire dataset, the WhARIO features separated low- and high-risk groups of patients with a hazard ratio of 2.29, whereas the tumor proportion score 1% reference threshold only reached a hazard ratio of 1.81. Combining the two yielded a higher hazard ratio of 2.60. See the WhARIO survival analysis framework for details.

The practical implication is that whole slide image features can complement existing biomarkers. When you develop a predictive model for clinical outcomes, consider whether combining image-derived features with established biomarkers improves performance beyond either alone. Document the hazard ratios and confidence intervals for each model so clinicians can assess the incremental value.

Data Management and Reproducibility

FAIR Principles for Image Data

The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable. These principles apply to whole slide image data as much as to genomic data. See the FAIR Guiding Principles publication for the foundational definitions.

For whole slide image analysis, applying FAIR principles means:

• Findable: Assign persistent identifiers to your whole slide images and metadata. Use standardized naming conventions that include the case identifier, tissue type, stain, and scan date.
• Accessible: Store images in a repository that supports controlled access. The NCBI Data Resources provide infrastructure for biomedical data sharing, and the EMBL-EBI Training resources cover data deposition and access.
• Interoperable: Use standard file formats and metadata schemas. Document the file format, compression, and color space for each image.
• Reusable: Provide detailed provenance information, including the scanner model, scanning settings, preprocessing steps, and analysis pipeline versions.

Storage and Cost Considerations

File size is a major expense in storage and data management for digital pathology workflows. The scanner comparison study found that scan area directly influences file size, and variations in scan area can profoundly impact operational efficiency. See the scanner tissue detection comparison for the relationship between scan area and file size.

For your own workflow, estimate storage requirements before you begin scanning. A single whole slide image can be several gigabytes when unpacked, and a research project with hundreds of slides will require terabytes of storage. Consider the following decisions:

• Store compressed file formats to reduce storage costs, but verify that compression does not degrade image quality for your analysis task.
• Use tiered storage, with active analysis data on high-performance storage and completed project data on archival storage.
• Establish a data retention policy that specifies how long raw images, processed tiles, and analysis outputs are retained.
• Back up critical data according to your institution's data management policy.

Data Sharing and Privacy

Data sharing raises privacy concerns, particularly for clinical data. The NIH Genomic Data Sharing Policy provides a framework for sharing genomic data responsibly, and similar principles apply to pathology image data. When you share whole slide images, you must consider whether the images contain identifiable patient information and whether your sharing plan complies with applicable regulations.

For research workflows, de-identify whole slide images before sharing by removing patient identifiers from filenames and metadata. For clinical workflows, follow your institution's policies for handling protected health information. Document your data sharing plan in your research protocol or data management plan.

Implementation in Routine Pathology Practice

Regulatory and Reimbursement Context

The implementation of digital pathology and artificial intelligence in routine pathology practice is shaped by regulatory frameworks. Several Food and Drug Administration-approved digital scanners and image management systems are establishing a foundation for the integration of advanced technologies into everyday pathology workflows. The regulatory environment, shaped by standards from the Food and Drug Administration, Centers for Medicare and Medicaid Services and Clinical Laboratory Improvement Amendments, and the College of American Pathologists, is evolving to accommodate these innovations while ensuring safety and reliability. Centers for Medicare and Medicaid Services and Clinical Laboratory Improvement Amendments have issued policies to allow pathologists to review and render diagnoses using digital pathology remotely. The introduction of new digital pathology Current Procedural Terminology codes designed to complement existing pathology Current Procedural Terminology codes is facilitating reimbursement processes. See the implementation of digital pathology and AI review for the full regulatory discussion.

For laboratories considering digital pathology implementation, the practical steps include:

1. Verify that your scanner and image management system meet applicable regulatory requirements.
2. Validate that pathologists can diagnose from digital slides with concordance to glass slide review.
3. Document the validation process, including the number of cases reviewed and the concordance rate.
4. Establish procedures for handling discrepant cases where digital and glass slide diagnoses differ.
5. Train staff on the digital pathology workflow and document their competency.

Quality-Controllable Report Generation

Recent methods for pathology report generation from whole slide images can produce slide-level diagnostic descriptions but often fail to ground fine-grained statements in localized visual evidence. A framework called QCAgent addresses this limitation by incorporating a customized critique mechanism guided by a user-defined checklist specifying required diagnostic details and constraints. The framework re-identifies informative regions in the whole slide image based on critique feedback and text-patch semantic retrieval, a process that iteratively enriches and reconciles the report. By making report requirements explicitly prompt-defined, constraint-aware, and verifiable through evidence-grounded refinement, QCAgent enables controllable generation of clinically meaningful and high-coverage pathology reports from whole slide images. See the QCAgent framework for pathology report generation for details.

For your own reporting workflow, consider whether automated report generation can support or augment pathologist review. Automated reports should be treated as drafts that require pathologist verification, not as final diagnoses. Document the checklist used to guide report generation and the process for verifying that each diagnostic detail is grounded in localized visual evidence.

Common Failure Patterns and How to Avoid Them

Data Leakage in Model Validation

Data leakage occurs when information from the test set influences training decisions. In whole slide image analysis, leakage can occur when tiles from the same slide appear in both training and test sets, or when slides from the same patient appear in both sets. The CAMELYON16 study avoided this by providing a separate training set and an independent test set, but many published studies do not follow this rigor. See the CAMELYON16 algorithm assessment for the validation methodology.

To avoid data leakage in your own workflow, split data at the patient or case level, not the slide or tile level. Verify that no patient appears in both training and test sets. Document your splitting strategy so reviewers can assess the validity of your performance estimates.

Overfitting to Scanner or Staining Artifacts

Models trained on slides from a single scanner or staining batch may learn scanner-specific or stain-specific features instead of biological features. This leads to poor generalization when the model is applied to slides from a different scanner or laboratory. The foundation model review notes that task-specific models often struggle to generalize to new datasets or unseen variations in image acquisition, staining techniques, or tissue types. See the foundation models review for the generalization discussion.

To avoid overfitting to technical artifacts, train and evaluate your model on data from multiple scanners and staining batches. If you cannot obtain multi-site data, use color normalization and artifact detection to reduce technical variation. Document the scanner and staining information for each slide in your dataset so you can assess whether your model generalizes across technical conditions.

Ignoring Image Quality in Downstream Analysis

Poor image quality can corrupt downstream analysis even when the model itself is well-designed. The GuardQC study emphasizes that reliable whole slide image analysis depends on image quality control that can be assessed before models are applied. See the GuardQC artifact quality control study for the quality control methodology.

Incorporate image quality assessment into your workflow as a gate before model application. If a slide fails quality control, either rescan the slide or exclude it from analysis. Document the quality control results for every slide so you can assess whether quality issues bias your results.

Limitations and Interpretation Boundaries

Two-Dimensional Analysis Limitations

Most pathology image analyses are limited to two-dimensional datasets, resulting in substantial information loss and incomplete interpretation. A framework for three-dimensional whole slide image analysis was developed and demonstrated on 3D vessel structure analysis with liver tissue sections. The workflow includes components on image registration, vessel segmentation, vessel cross-section association, object interpolation, and volumetric rendering. See the 3D vessel reconstruction framework for details.

For your own analysis, recognize that two-dimensional whole slide images capture only a single section through a three-dimensional tissue sample. Measurements of cell density, vessel diameter, or fibrosis extent may not fully represent the three-dimensional structure. When you interpret your results, acknowledge this limitation and consider whether three-dimensional reconstruction is necessary for your research question.

Model Performance Depends on Data Quality

The performance of any computational model depends on the quality and representativeness of the training data. The CAMELYON16 study found that algorithm performance ranged from 0.556 to 0.994 AUC, a wide range that reflects differences in model architecture, training data, and preprocessing. See the CAMELYON16 algorithm assessment for the performance range.

When you evaluate a model, consider whether the training data represents the population and conditions where you plan to apply the model. A model trained on slides from two centers in the Netherlands may not generalize to slides from other countries with different staining protocols and patient populations. Document the training data provenance and assess whether your application context matches the training context.

Quantitative Outputs Require Clinical Interpretation

Quantitative outputs from whole slide image analysis, such as fibrosis percentage or cell counts, require clinical interpretation. A hybrid model for quantifying kidney interstitial fibrosis from trichrome-stained whole slide images demonstrated strong agreement with pathologist-derived estimates at both the biopsy core and patient levels. The R2 values for the validation set was 0.86 and increased to 0.93 on the subset of validation cores with good image quality, and the mean difference between predicted and ground-truth fibrosis estimates was always less than 3%. See the TRI_IF kidney fibrosis quantification model for details.

The model's performance depended on image quality, with better agreement on cores with good image quality. This reinforces the importance of quality control before model application. When you report quantitative outputs, include the confidence intervals and note the image quality of the underlying slides.

Professional Escalation Criteria

When to Escalate to Pathologist Review

Computational models can support but should not replace pathologist review for diagnostic decisions. The All Wales study found 86 clinically significant discrepancies between digital and glass slide reports, with ground truth lying with digital pathology in 28 cases and with glass slide in 58 cases. See the All Wales digital pathology verification study for the discrepancy analysis.

Escalate to pathologist review when:

• The model output conflicts with clinical information or prior diagnoses.
• The model produces a low confidence score or ambiguous classification.
• The slide fails quality control or contains artifacts that may affect model performance.
• The model output would change patient management if acted upon without verification.
• The case involves a rare or unusual presentation not well represented in the training data.

Documentation for Discrepant Cases

Document discrepant cases systematically to support quality improvement and regulatory compliance. For each discrepant case, record:

• The model output and confidence score.
• The pathologist's diagnosis from digital review.
• The pathologist's diagnosis from glass slide review, if performed.
• The ground truth determination and the method used to establish it.
• The clinical significance of the discrepancy.

The All Wales study used a panel of expert pathologists to review discrepant cases under a multiheader microscope to establish ground truth. See the All Wales digital pathology verification study for the ground truth methodology. For your own workflow, establish a similar process for resolving discrepancies and document the outcomes.

Frequently Asked Questions

What is the difference between whole slide imaging and digital pathology?

Whole slide imaging is the process of scanning glass slides to create digital images. Digital pathology is the broader field that encompasses whole slide imaging, image management, computational analysis, and the use of digital images for diagnosis, research, and education. Whole slide imaging allows pathologists to view and analyze microscopic images on a computer monitor, enabling computational pathology. See the foundation models review for the transformation of traditional pathology into a computer vision workflow.

How large are whole slide image files?

Whole slide image files are typically several gigabytes when unpacked. This large file size forces images to be tiled into a series of subsamples for modeling. See the pre- and post-processing review for WSI analysis pipelines for the file size discussion. The actual file size depends on the scan area, magnification, compression, and file format. Scanner tissue detection algorithms that include more non-tissue area in the scan region produce larger files.

What is tiling in whole slide image analysis?

Tiling is the process of dividing a whole slide image into smaller subsamples or tiles that can be processed by computational models. Because whole slide images are too large to process in a single pass, models analyze a series of tiles and the results are aggregated. See the pre- and post-processing review for WSI analysis pipelines for the tiling discussion. Tile size, overlap, and sampling strategy affect model performance and computational cost.

How do I choose between a task-specific model and a foundation model?

Task-specific models, such as convolutional neural networks, perform well when you have a large annotated dataset for a specific task. Foundation models are large AI models trained on sizable data and adapted or fine-tuned to perform different tasks using a modest amount of task-specific annotated data. See the foundation models review for the comparison. Choose a task-specific model when you have abundant annotations and a well-defined task. Choose a foundation model when you have limited annotations or need to generalize across staining variations and tissue types.

What quality checks should I perform before running a model on whole slide images?

Before running a model, verify that the scanner produces adequate image quality, check for artifacts that may affect model performance, and confirm that the tissue detection algorithm captured the relevant tissue. The GuardQC study demonstrates a conservative preprocessing pipeline for artifact control and tissue preservation. See the GuardQC artifact quality control study for the quality control methodology. Establish thresholds for acceptable artifact levels and masked tissue percentage based on your tissue type and analysis task.

How do I validate a whole slide image analysis model?

Validate your model on an independent test set that was not used for any training decisions. Compare your model's performance against a relevant baseline, such as a pathologist's diagnosis or a published benchmark. Report confidence intervals for your performance metrics. The CAMELYON16 study provides a model for validation, with a training set from two centers and an independent test set evaluated by both algorithms and pathologists. See the CAMELYON16 algorithm assessment for the validation methodology.

What are the main challenges in whole slide image analysis?

The main challenges include managing large file sizes, handling imbalanced datasets, generalizing across staining and scanner variations, and interpreting quantitative outputs in a clinical context. Task-specific models often struggle to generalize to new datasets or unseen variations in image acquisition, staining techniques, or tissue types. See the foundation models review for the generalization challenges. Sample imbalance remains an intractable challenge for multiclassification tasks. See the cross-patient pseudo bags generation method for the imbalance discussion.

How do I handle discrepant cases between digital and glass slide review?

Establish a process for resolving discrepancies before you implement digital pathology. The All Wales study used a panel of expert pathologists to review discrepant cases under a multiheader microscope to establish ground truth. See the All Wales digital pathology verification study for the discrepancy resolution methodology. Document each discrepant case, the ground truth determination, and the clinical significance of the discrepancy. Escalate to pathologist review when the model output conflicts with clinical information or would change patient management.

Related Bioinformatics Guides

• Ethical Considerations in Computational Genomics
• Circular RNAs: Computational Identification and Analysis
• Alternative Splicing Analysis from RNA-Seq Data
• Pan-Cancer Analysis of Whole Genomes (PCAWG)
• Computational Analysis of Viral Protease Inhibitors and Drug Resistance

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.
• WhARIO: whole-slide-image-based survival analysis for patients treated with immunotherapy.. Journal of medical imaging (Bellingham, Wash.), 2024.
• Diagnostic Assessment of Deep Learning Algorithms for Detection of Lymph Node Metastases in Women With Breast Cancer.. JAMA, 2017.
• Whole-slide imaging and a Fiji-based image analysis workflow of immunohistochemistry staining of pancreatic islets.. MethodsX, 2022.
• Revolutionizing Digital Pathology With the Power of Generative Artificial Intelligence and Foundation Models.. Laboratory investigation, a journal of technical methods and pathology, 2023.
• Implementation of Digital Pathology and Artificial Intelligence in Routine Pathology Practice.. Laboratory investigation, a journal of technical methods and pathology, 2024.
• Introduction to Digital Image Analysis in Whole-slide Imaging: A White Paper from the Digital Pathology Association.. Journal of pathology informatics, 2019.
• LIVER WHOLE SLIDE IMAGE ANALYSIS FOR 3D VESSEL RECONSTRUCTION.. Proceedings. IEEE International Symposium on Biomedical Imaging, 2015.
• Verification and Validation of Digital Pathology (Whole Slide Imaging) for Primary Histopathological Diagnosis: All Wales Experience.. 2021.
• Erratum regarding missing Declaration of Competing Interest statements in previously published articles. 2026.
• Tissue-Preserving Artifact Quality Control for Whole-Slide Computational Pathology.. 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.
• Classifying driver mutations of papillary thyroid carcinoma on whole slide image: an automated workflow applying deep convolutional neural network. Frontiers in Endocrinology, 2024.
• An Open-Source Whole Slide Image Registration Workflow at Cellular Precision Using Fiji, QuPath and Elastix. Frontiers of Computer Science, 2022.
• Cross-Patient Pseudo Bags Generation and Curriculum Contrastive Learning for Imbalanced Multiclassification of Whole Slide Image. arXiv.org, 2024.
• QCAgent: An agentic framework for quality-controllable pathology report generation from whole slide image. arXiv.org, 2026.
• A deep learning and morphometric hybrid model for automated quantification of kidney interstitial fibrosis in trichrome-stained whole-slide image. Journal of Pathology Informatics, 2026.

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