Computational pathology is increasingly mentioned in tumor boards, journal clubs and procurement meetings, yet the vocabulary can feel borrowed from computer science. This article is for pathologists, oncologists and other clinicians who want a working understanding of the core concepts, enough to read a paper or vendor sheet critically and ask the right questions.
What computational pathology means
Computational pathology is the use of computational methods to extract information from pathology data, most often digitized slides, sometimes combined with clinical, genomic or report data. It includes classical image analysis built on hand-crafted rules as well as machine learning, where the system learns patterns from examples.
It is broader than AI, and broader than diagnosis. Measuring tumor area, quantifying stain intensity, searching an archive for similar cases and building research cohorts all fall within its scope. For a plain overview of where these tools stand clinically, see what AI in pathology can and cannot do today.
From glass to data: the whole slide image
Everything starts with a whole slide image, a digital capture of an entire glass slide at high magnification. These images are very large, typically far larger than a radiology study, and they are stored as a pyramid of resolutions so a viewer can zoom smoothly. If you want the technical background, read how whole slide imaging turns glass into a gigapixel image.
Because a full image is too large to feed into most models at once, it is usually divided into small square tiles, also called patches. A model analyzes tiles, and the results are stitched back together into a map or summarized into a slide-level answer. This has consequences:
- The model may only see a small field at a time, so architectural context across the slide must be handled explicitly.
- Background, pen marks and out-of-focus areas must be filtered out first, or they add noise.
- Magnification matters. A model trained at one magnification can behave differently at another.
- A slide shows only the tissue that was sectioned and scanned, so a slide-level result reflects that sample, not the whole block or lesion.
Core concepts in computational pathology
Tasks: what is the model asked to do?
Most tools perform one of a few task types:
- Classification assigns a label to a tile, region or slide, such as tumor versus non-tumor.
- Detection finds and locates objects, such as mitotic figures or nuclei.
- Segmentation outlines regions pixel by pixel, such as tumor boundaries or glands.
- Regression or scoring predicts a number, such as a proportion of positive cells.
Knowing the task tells you how output should be checked. A segmentation can be judged by looking at the outline. A slide-level classification needs more care, because you cannot see directly why the model decided.
Training data and labels
A machine learning model is only as good as the examples it learned from. Labels might be detailed pixel-level annotations drawn by pathologists, or weaker slide-level labels taken from signed reports. Detailed labels are expensive to produce; weak labels are plentiful but noisier. Ask who labeled the data, how disagreements were resolved, and how many institutions, scanners and staining protocols were represented.
Ground truth is not always truth
In pathology, the reference standard is often a pathologist's diagnosis, which carries its own inter-observer variability. A model that agrees closely with one expert panel may simply have learned that panel's habits. Where possible, stronger reference standards include consensus reads, follow-up outcomes or molecular results.
Performance metrics
Papers and brochures report sensitivity, specificity, area under the ROC curve and similar figures. These summarize performance on a particular test set. They do not tell you how the model performs on your slides. Look for whether the test set came from institutions that were not used in training, and whether results are broken down by subgroup.
Also check whether the reported operating threshold is the one the product actually uses, since sensitivity and specificity trade off against each other as the threshold moves.
Rules versus learning
Older image analysis tools follow explicit rules set by developers, such as color thresholds and size limits for nuclei. They are transparent and predictable, but brittle when staining or tissue varies. Machine learning models learn their own features from examples, which makes them more flexible and often more accurate, but harder to inspect. Many practical products combine both, using rules for steps like tissue detection and learned models for the interpretive part. Knowing which is which helps you anticipate failure: rules tend to break visibly when conditions change, while learned models can drift quietly.
Where it can go wrong
Clinicians should know the typical failure modes, because they are not always obvious in the output.
- Domain shift. Differences in staining, scanners, tissue processing or populations can reduce accuracy without any warning.
- Shortcut learning. Models sometimes key on irrelevant signals, such as a site-specific stain shade or scanner artifact, that happened to correlate with the label.
- Class imbalance. Rare diagnoses are underrepresented, so performance on them is uncertain.
- Overconfidence. A model can output a high probability for an image unlike anything it was trained on.
None of this makes computational pathology unreliable as a field. It means each tool needs local verification and sensible guardrails, just like any new laboratory method.
What clinicians should ask
When a computational result appears in a report or a meeting, a few questions help put it in context:
- What exactly was measured or predicted, and at what level: tile, region, slide or patient?
- Was the tool validated in this laboratory, on this scanner and staining protocol?
- Did a pathologist review the output before it reached the report?
- How does the result change management, if at all, compared with conventional assessment?
- What is known about performance in patients like this one?
Pathologists remain responsible for the interpretation. A computational score is evidence to weigh, not a verdict.
Key takeaways
- Computational pathology extracts information from digitized slides and related data, using both rule-based analysis and machine learning.
- Whole slide images are processed in tiles, so magnification, tissue filtering and slide-level context all affect results.
- The quality and diversity of training labels largely determine how well a model generalizes.
- Published metrics describe a test set, not your lab; local verification is essential.
- Clinicians should ask what was measured, how it was validated and who reviewed it.
As labs build digital workflows, the groundwork matters as much as the algorithms. PATHOLOGY+ supports whole slide image viewing and collaboration alongside the case and report workflow. Book a demo to see it with your own team.