AI Study Shows Cancer Therapy Effectiveness

A new study from researchers at UCL suggests that an AI-assisted way of reading tumour biopsies could sharpen who benefits from an intensified therapy for locally advanced rectal cancer. The work, published in eBioMedicine, examines whether adding irinotecan to standard chemoradiotherapy improves survival—specifically for patients whose tumours show a high density of cancer cells at diagnosis.

Rather than treating irinotecan as universally helpful, the researchers used artificial intelligence to stratify patients based on the microscopic composition of their biopsy samples. Their model separates tumour samples into “high” and “low” cancer-cell density groups by analysing standard histology images, enabling a pattern that conventional assessment had not captured reliably at scale.

This distinction matters because the ARISTOTLE trial, which enrolled patients across 75 UK hospitals, previously reported limited overall benefit from adding irinotecan. Here, the same dataset is re-analysed through the lens of tumour biology quantified by AI, revealing a subgroup where the intensified regimen performs substantially better.

For patients classified with high tumour cell density, adding irinotecan reduced the risk of cancer recurrence by about 43% and cut the risk of death by about 50% over five-year follow-up compared with the typical treatment arm using capecitabine plus radiation. In contrast, patients classified as low density showed no clear survival advantage from the intensified approach.

The technical core of the study is the AI training pipeline. The researchers trained their system on large open-source datasets and then applied it to microscopic slides from 414 patients in the phase III trial. The model both recognises tumour tissue and performs cell-level classification—counting cancerous versus healthy cells—more efficiently than manual annotation.

By automating this “cell density” measurement, the approach makes it feasible to process hundreds of biopsies with consistent criteria. The result is a clinically relevant biomarker-like signal: a quantitative readout tied to treatment response rather than a subjective visual impression.

To support uptake, the team also built a free online tool called Octopath, allowing clinicians to upload biopsy slides for analysis. While promising, the authors emphasise that independent verification and further prospective clinical evaluation are needed before the method can guide routine treatment decisions or broaden access to irinotecan-based therapy.

If validated, the study points toward a viral-looking new paradigm for oncology: pairing existing therapies with AI-derived tumour phenotyping to deliver precision intensification—helping the right patients while sparing others from avoidable toxicity.