AI Doesn’t Just Examine Pathology Slides—It Has Also Learned How Experts Search for Lesions

A whole-slide digital pathology image may contain billions of pixels, yet the abnormalities that truly determine a diagnosis may occupy only a tiny corner. Pathologists do not examine every inch of tissue equally. Instead, they scan structures at low magnification, identify suspicious regions, and then zoom in to examine cells. A study published in *Nature Biomedical Engineering* transformed this search process—which is generally not documented in diagnostic reports—into data that artificial intelligence can learn from.

The research team proposed the Pathology-CoT framework, using an “AI Session Recorder” to capture pathologists’ panning, zooming, and dwell behavior in a slide viewer, then organizing the continuous actions into standardized instructions and bounding boxes. Publicly available information indicates that the training material came from 10.6 hours of diagnostic sessions conducted by eight Stanford School of Medicine pathologists, producing 5,222 rounds of behavioral and reasoning dialogue.

The approach not only identifies “where to look,” but also attempts to explain “why to look there.” The system first drafts the rationale for each observation step, which pathologists then review and revise, creating supervised data that pairs locations with diagnostic significance. The research team estimates that this human-supervised process can increase annotation speed by approximately sixfold, although its quality still depends on whether experts fully correct the AI-generated explanations.

Pathology-o3, trained on these data, uses a two-stage design: it first predicts regions of interest worth examining, and then has a vision-language model analyze them sequentially. In internal testing for gastrointestinal cancer lymph node metastasis, the preprint reported precision of 84.5%, recall of 100.0%, and accuracy of 75.4%, outperforming several vision-language models included in the study’s comparisons. The three metrics measure different aspects of performance; 100% recall means that no positive cases were missed in that test set, not that every judgment was correct.

The model still achieved 97.6% recall on the independent external LNCO2 dataset, indicating that its ability to locate metastatic lesions did not disappear entirely when applied to a different set of slides. The journal paper also reported that its benefits could be reproduced across different vision-language model backbones, and separately used dermatopathology data to evaluate region localization and reasoning. However, the current results primarily demonstrate the feasibility of a training strategy and do not yet show that the system can handle the wide range of cancer types, benign changes, and rare pitfalls encountered in routine pathology.

More practical challenges remain before clinical deployment: differences among laboratories in staining, scanners, and tissue preparation methods may alter images, while high recall may also be accompanied by more false positives, increasing physicians’ review burden. Future research must evaluate diagnostic errors, slide-reading time, human–machine responsibilities, and regulatory requirements across more institutions and in prospective workflows. The core value of this work, therefore, is not to have AI imitate an elegant piece of written reasoning, but to turn the way experts progressively find evidence within enormous slides into a testable and reproducible learning objective.