Reading Recurrence Risk at the Tumor Border: AI Helps Identify Stage II Colorectal Cancer

After surgery for stage II colorectal cancer, the truly difficult question is often not whether the tumor has been removed, but who may still experience recurrence. If risk is underestimated, patients may miss out on close monitoring or adjuvant therapy; if it is overestimated, they may endure chemotherapy side effects that are not necessarily warranted. An Australian research team is now attempting to use artificial intelligence to provide another measure of risk based on existing pathology data.

The deep learning system, called SÉMIL, analyzes both digitized pathology slides and their textual descriptions to classify stage II patients as having higher or lower recurrence risk. Rather than searching for entirely new molecular markers, the study reinterprets slides that hospitals already routinely produce, focusing particularly on the “invasive front,” where the tumor advances into surrounding tissue. Growth patterns in this area have prognostic significance, but their morphological complexity may make consistent classification by different pathologists difficult.

The researchers analyzed more than 1,600 pathology slides and validated the system in three independent patient cohorts from multiple Australian institutions, covering a total of 1,220 patients with stage II colorectal cancer. The study was published in *Gastroenterology* and led by Francis Magisson of La Trobe University, with collaborating institutions including Austin Health, the Olivia Newton-John Cancer Research Institute, WEHI, Monash University, and the University of Melbourne.

The results showed that risk stratification was most accurate when SÉMIL’s classification agreed with the pathologist’s interpretation. This also defines a more realistic role for the tool: it is not intended to replace pathologists, but to strengthen assessment in areas prone to disagreement by providing a second method of interpretation. If subsequent research demonstrates clinical benefit, higher-risk patients could potentially receive more intensive follow-up and be evaluated for postoperative chemotherapy by their healthcare teams; however, the model’s classification alone does not mean it has been proven that a patient will benefit from chemotherapy.

SÉMIL’s practical appeal is that it requires neither additional tissue collection nor the prior adoption of expensive new tests and could theoretically be integrated into existing digital pathology workflows. However, currently available university and news materials do not provide complete performance figures such as sensitivity, specificity, or discrimination curves, and are insufficient to determine the model’s stability across different scanning equipment, staining workflows, and patient populations.

Therefore, although independent validation across institutions is more robust than testing at a single hospital, it is still not the endpoint for clinical implementation. The research team next needs to confirm through prospective studies whether adding SÉMIL genuinely improves physicians’ risk assessments, leads to more appropriate treatment choices, and benefits patients’ recurrence and survival outcomes. Until these questions are answered, it is better suited as a decision-support tool than as a standalone basis for determining monitoring or chemotherapy.