AI interpretation of tissue slides

Digitise a tissue section faithfully and you hold roughly a hundred thousand by a hundred thousand pixels — gigapixel territory, orders beyond what image networks digest whole. Computational pathology is the discipline built around that mismatch: it succeeds exactly where it reorganises the question into pieces learnable from data that actually exists. Two inventions carry almost everything.

Bags of tiles, verdicts about bags

The workable reformulation slices a slide into thousands of small patches and treats the slide as one labelled item — cancer or clear, high grade or low. Which particular tile carries the evidence stays unknown during training; only the aggregate verdict is trustworthy. Multiple-instance learning formalises this bargain: a model scores every patch, an attention layer learns which patches probably justified the label, and the slide-level decision aggregates their testimony. The payoff exceeds classification — attention maps double as heatmaps flagging suspicious foci for human review, converting the network from replacement into search instrument. One cultural benefit deserves note: forcing the label structure into the open exposed how much diagnostic teaching had always relied on slides pointing at regions no one ever delineated rigorously.

Learning before anyone labels

Expert annotation is the scarcest resource in medicine, so modern systems pretrain without it: self-supervised encoders learn visual grammar from millions of unlabelled tiles by predicting hidden relationships within their own pixels, then absorb task supervision afterwards from far smaller labelled sets. This division — unsupervised structure, supervised judgement — repeats everywhere deep learning entered medicine, but pathology set its cleanest example because raw imagery was abundant while ground truth was rationed.

Search, judgement, and the surrogate frontier

Task selection determines reliability more than architecture does. Detection work suits these systems superbly: sweeping lymph-node levels for metastatic deposits or prostate cores for tiny acinar foci is exhaustive-search behaviour, immune to fatigue, and error analysis becomes asymmetric — missed tiles matter, flagged-but-negative tiles merely cost seconds. Grading sits harder, because tumour-grade categories were shaped by consensus among reading humans whose own repeat-agreement limits form the true baseline; measured against that baseline, algorithmic concordance proves respectable rather than miraculous. The seductive frontier predicts molecules from appearance — microsatellite instability, mutation class, even expression surrogates visible in molecular profiles otherwise demanding sequencing — exploiting real biological coupling between how cells look and how their genomes misbehave. Every such predictor is however a statistical shadow with performance ceilings near, not above, direct measurement, best treated as triage deciding who proceeds to definitive testing.

Two structural caveats complete an honest picture. Domain shift bites imaging as hard as genomics: models tuned on one scanner’s colour response and one laboratory’s staining ritual stumble on another, so normalisation and per-site validation become permanent obligations, monitoring drift across instruments and years like assay quality control. And responsibility never transferred: deployed products flag, measure and prioritise while a licensed pathologist signs, which keeps today’s systems adjuncts by law rather than by limitation — a status reflecting both current accuracy and the profession’s refusal to orphan accountability.