Can a Single Slide Determine Treatment Intensity? AI Identifies a Group That May Benefit in a Rectal Cancer Trial

Locally advanced rectal cancer presents a difficult treatment boundary: intensifying chemotherapy may reduce the risk of recurrence, but it may also add burdens such as severe diarrhoea and reduced white blood cell counts. A UK research team is now attempting to use pathology slides already obtained at diagnosis to identify patients who are genuinely likely to benefit from intensified treatment, so that treatment intensity is no longer determined solely by broad risk classifications.

Researchers used an artificial intelligence pathology model to analyze 414 rectal cancer biopsy slides from the phase 3 ARISTOTLE trial. The model first identified the tumour and surrounding tissue, then calculated the density of cancer cells within them. Of the 414 patients, 188 were classified as having high tumour cell density and 226 as having low density. The method processes routine microscopic images and does not require additional genetic sequencing or special staining.

ARISTOTLE originally compared two preoperative regimens: capecitabine combined with radiotherapy, and the same regimen with irinotecan added. The overall trial had previously shown only limited benefit from adding irinotecan; however, in this post hoc analysis, patients with high tumour cell density who received the intensified regimen had an approximately 43% lower risk of cancer recurrence and an approximately 50% lower risk of death. No clear difference was observed among patients with low density.

The significance of this result is not that AI independently selected a new drug, but that it converted a signal in pathology slides that is difficult to quantify rapidly and consistently by hand into a biomarker that may predict treatment benefit. If subsequently confirmed to be effective, clinicians may be able to identify before treatment begins those who are more likely to benefit from irinotecan, while allowing other patients to avoid unnecessary toxicity.

However, this remains a retrospective analysis seeking differences in treatment response within existing trial data, rather than a clinical trial that was prospectively designed specifically to validate an AI-based subgrouping strategy. The sample also included only the 414 slides from the original trial that were available for analysis. Existing data cannot yet adequately determine whether differences among hospitals in sampling, staining, scanning equipment, and image quality would alter the classification results.

Therefore, the approximately 43% and 50% figures should be regarded as estimates of association within a specific subgroup and cannot be directly applied to all patients with locally advanced rectal cancer. The research team has made the analytical tool into an online system to which slides can be uploaded, but until it has been validated in independent patient groups, prospectively planned studies, and clinical workflows, it cannot be used to determine whether an individual patient should receive added irinotecan. Model quality control, accountability, and medical software regulation are also issues that must be addressed before it enters routine care.