Digital pathology: distinguish a scan from an inference
In one sentence
Digital pathology converts prepared slides into images for viewing and analysis; a scanned image, a computational estimate and a validated clinical interpretation are different outputs.
The intuition
A high-quality photograph of a map lets someone inspect the map remotely. Software might then estimate building counts or classify neighborhoods. The photograph and the estimates are different products. Digital pathology follows that distinction: acquiring pixels preserves visible information, while interpreting those pixels requires another method. The analogy stops at image quality, because small missed cellular details can matter diagnostically.
How it works
Whole-slide imaging (WSI) uses a scanner to acquire a digital representation of the prepared glass slide. The workflow includes optics, focus, color capture, file handling, display and a viewing application. “Whole slide” does not mean every cell in the original tissue block or body; the image covers the scanned section and captured areas.
A pathologist can navigate the image and interpret morphology within a validated diagnostic workflow. A computational method may instead segment objects, meaning mark their boundaries, count selected cells, estimate areas or generate a model score. Its output depends on the algorithm's defined task and the images on which it was developed and evaluated.
A further model might predict molecular features from appearance. Such a prediction is not a direct RNA or protein assay. The model needs reference measurements and testing in suitable independent data to support its claim. Reliable viewing of a scanned slide does not validate every algorithm run on that image.
Why it matters in cancer
Images can support remote review, teaching and repeated analysis without cutting another section from the block, when the required slide already exists. This helps preserve finite tissue. Nevertheless, an existing image may omit another depth, stain or region needed for a new question.
An algorithm trained on one set of stains, scanners or populations may behave differently elsewhere. Generalizability asks whether performance carries into the intended new setting. A convincing overlay or a precise-looking decimal does not settle that question. The FDA digital pathology research program explicitly distinguishes technical image performance from clinical performance and identifies generalizability as a research need.
How it is measured
| Assay-card field | What to retain |
|---|---|
| Measures | Visible slide information; a separate algorithm measures or estimates a defined feature |
| How | Scan → quality check → view or run defined analysis → interpret within its evidence scope |
| Input and tissue cost | A prepared slide; scanning itself generally uses no additional tissue section |
| Output and units | Image pixels/resolution, selected cell counts or areas, or named model scores with their definitions |
| Thresholds | Quality criteria and model cutoffs depend on the scanner, task and intended use |
| Failure modes | Missed tissue, poor focus, folds, stain/color variation, compression or incorrect cell boundaries |
| Cannot show alone | Unscanned tissue, directly measured molecular expression or validated treatment benefit |
| Validation context | Diagnostic WSI validation covers the intended workflow; each computational clinical claim requires suitable additional evidence |
Common confusions
- A digital slide is not an automatically generated diagnosis.
- A validated scanner workflow does not validate every downstream artificial-intelligence model.
- A heat map marks a method's output; its biological interpretation depends on the method.
- Reanalysis can reuse pixels, but it cannot reveal a tissue region that was never scanned.
Try it
A fictional H&E scan passes image-quality checks. Research software labels a region “high predicted RNA signal.” A learner reports that the gene was measured at high expression in that region. What changed improperly?
Answer: A prediction became a direct measurement. The result should remain a research estimate with its model, reference data and validation scope. Measuring RNA would require an appropriate molecular assay; the good scan does not supply that assay.
Explain it back
What are the three layers to retain? One answer: The acquired image, the method that interprets or estimates from it, and evidence that the resulting interpretation is valid for its intended use.
Takeaway
Keep pixels, computational estimates and clinical interpretations separate, even when they appear in one viewer.
Related concepts
Sources and scope
Source check: October 10, 2026. General digital measurement teaching; expert and learner review remain pending.
- CAP: validating WSI for diagnostic purposes, 2021 update — intended clinical applications and diagnostic-workflow validation; published in print in 2022.
- FDA: digital pathology research program — technical performance, image quality, computational metrics and generalizability boundaries.