AI digital pathology connects the cancer slide with treatment decisions

Digital pathology is opening new routes to faster diagnosis and more selective cancer treatment. European implementation studies reveal where AI already changes laboratory work and where precision oncology still needs stronger validation.

A stained tissue section contains more information than a cancer label. Its cells, architecture and surrounding tissue help pathologists establish what disease is present and how it should be classified. Artificial intelligence is extending the questions researchers ask of those images, including whether they can indicate molecular characteristics or identify patients most likely to benefit from a treatment.

These ambitions involve different decisions. Highlighting a small suspicious focus for review is not equivalent to recommending hormone therapy. Nor does a model that forecasts an unfavourable prognosis necessarily identify the treatment that will improve it. The clinical value of digital pathology depends on connecting each output to the particular decision it can support.

For European laboratories, the immediate opportunity may be relatively practical: fewer unnecessary stains, shorter waits for authorised reports and better support for difficult cases. Alongside this, research is testing whether routine slides can carry some of the information currently obtained through additional molecular assays. Together, these developments could change the relationship between the pathology laboratory and the cancer multidisciplinary team.

UK implementation reveals differences between sites

The Articulate Pro study, published in May 2026, evaluated AI assistance in prostate biopsy reporting across three NHS specialist centres in England. It included 1,613 cases, of which 1,049 were reported with assistance. In a staged second reading phase, AI prompted changes to the initial diagnosis or grade group in 21 of 386 patients; five changes potentially affected clinical management.

The workflow results were instructive. Concurrent AI use was associated with an adjusted reduction of 30.1 hours in mean turnaround time at one site. The other two sites did not show statistically significant turnaround improvements. Requests for immunohistochemistry, which uses additional stains to investigate tissue characteristics, fell significantly at all three.

This observational introduction into routine practice does not establish that every laboratory would obtain the same benefit. It also separates a changed pathology report from a demonstrated improvement in a patient’s eventual outcome.

The variation matters to procurement. A laboratory should establish which delay it is trying to solve and whether software can influence that stage. An average gain from another hospital is a starting hypothesis for local service evaluation, rather than a guaranteed delivery target.

Fewer stains need not mean faster reading

The Dutch CONFIDENT P study provides a smaller but detailed test of selective additional testing. Published in 2025, it allocated patients in alternating two week periods to standard assessment or an AI assisted workflow. Across 82 patients and 237 slides, assistance reduced immunohistochemistry use per detected prostate cancer case, with a patient level relative risk of 0.55.

However, median assessment time per routine stained slide was not significantly different: 139 seconds with assistance and 112 seconds in the control arm. The intervention did not simply eliminate confirmatory investigation. Immunohistochemistry was requested for all slides considered morphologically negative in the AI arm.

That detail prevents an important misinterpretation. The study supports a particular testing pathway with retained safeguards. It is not evidence that a reassuring AI classification makes additional testing unnecessary in every uncertain biopsy.

Laboratories assessing the commercial case should therefore separate slide reading, ancillary testing and the total time before a report is authorised. These can move in different directions. A system may add a review step yet reduce the need for another laboratory procedure; whether that helps patients depends on the surrounding workflow.

Molecular screening from an ordinary slide

In colorectal cancer, researchers have investigated whether images can help identify microsatellite instability, a molecular feature associated with defects in DNA mismatch repair. The MSIntuit study, published in Nature Communications in 2023, blindly validated an AI prescreening tool using an independent dataset of 600 consecutive patients. Slides were digitised on two scanners.

Sensitivity was 96–98%, while specificity was 46–47%. In the tested setting, the system could rule out almost half of patients without the biomarker while retaining high sensitivity. This is a proposed way to reduce the population requiring subsequent testing, rather than a claim that the image directly measures every underlying molecular alteration.

The remaining false negatives are clinically important. Performance for uncommon patterns was uncertain because few cases were available. The study also incorporated image quality checks and calibration, which belong to the evaluated process rather than being optional extras.

For a service considering such an approach, the relevant balance includes tests avoided and eligible patients missed. Those consequences are asymmetric: saving laboratory work cannot compensate automatically for failing to identify a result that changes a person’s cancer assessment or treatment options.

Predicting treatment benefit requires a different test

A 2023 NEJM Evidence study addressed a more ambitious question in localised prostate cancer: who benefits from adding short term androgen deprivation therapy to radiotherapy? Researchers developed a model using pathology images and clinical information from randomised trial populations, then evaluated a locked model in 1,594 participants from NRG/RTOG 9408. AI model for hormone therapy benefit

Among the 34% classified as model positive, estimated distant metastasis at 15 years was 4.0% with added hormone therapy and 14.4% with radiotherapy alone. In the model negative group, estimates were 6.9% and 7.4%, without a statistically significant treatment difference.

The original treatment allocation was randomised, strengthening assessment of differential benefit. However, this was a retrospective biomarker analysis of trial material, not a prospective trial assigning care according to the AI recommendation. A nonsignificant difference in one subgroup does not guarantee that every person in it can safely omit treatment.

That distinction is central to precision oncology. A useful treatment biomarker must do more than separate patients with better and worse outlooks. It must support a choice between realistic alternatives, considering both cancer control and treatment burden. Prospective use also has to remain relevant to the treatments and patients seen today.

Kidney cancer exposes the problem of tumour variation

Another approach attempts to infer tumour blood vessel characteristics from routine histology. A 2025 Nature Communications study developed a deep learning score for clear cell renal cell carcinoma and examined its relationship with response to antiangiogenic treatment, which targets processes supporting blood vessel growth.

The research included a 145 patient clinical cohort and independent material from the IMmotion150 trial. In the trial material, the image derived score had a concordance index of 0.66 for progression free survival in patients receiving sunitinib, compared with 0.67 for the RNA based score it sought to approximate. A concordance index assesses how well predicted rankings align with observed outcomes; it is not a percentage of patients successfully treated.

The work suggests that tissue images can provide a practical route to studying a biological signal. It does not show that choosing treatment with the image model improves survival. The authors identify further clinical trial validation as necessary before adoption.

Tumour sampling creates another practical question. A single section may not represent every region of a heterogeneous tumour. An image based assay could make examining several regions more feasible, but its clinical interpretation must account for what tissue was sampled and how different areas contribute to the final score.

Broad models still need a defined clinical job

Foundation models seek to learn reusable image representations that support many pathology tasks. The Virchow study, published in Nature Medicine in 2024, reported a specimen level area under the receiver operating characteristic curve of 0.95 across nine common and seven rare cancers in a pan cancer detection evaluation. Its assessments also included biomarker prediction and cell identification.

Such breadth could reduce the work needed to develop individual applications. Yet an aggregate performance measure cannot specify the acceptable risk for every cancer, specimen type or intended use. A model used to prioritise review faces a different clinical test from one used to exclude a molecular investigation.

The most useful industrial development would connect this flexibility with clearly defined laboratory products. Services need to know which specimens are supported, when the system should decline to provide an answer and how its output enters an authorised report. Pathologists also need a workable means of resolving discrepancies without turning every case into a research exercise.

Digital pathology is therefore advancing along two connected tracks. Routine implementation is beginning to reveal how AI changes laboratory decisions and resources. Biomarker research is expanding what tissue images might contribute to treatment selection. Progress depends on preserving the distinction while building reliable links between them, so that additional information reaches the cancer team in a form that can improve a real decision.