An Australian-developed AI pathology tool could improve relapse prediction in stage II colorectal cancer, helping identify patients who may benefit from adjuvant chemotherapy using routine H&E slides.
An artificial intelligence tool developed by La Trobe University researchers has identified stage II colorectal cancer patients at significantly higher risk of relapse using routine pathology slides alone, potentially refining one of the most difficult treatment decisions in bowel cancer care.
The deep learning model, known as SÉMIL (Semantically-Enhanced Multiple Instance Learning), analyses standard haematoxylin and eosin-stained pathology slides to assess tumour growth patterns associated with recurrence risk.
In a multi-institutional validation study involving 1220 patients across three independent cohorts, the model consistently predicted five-year relapse-free survival and remained independently prognostic after adjustment for established clinicopathological risk factors.
The findings, published in Gastroenterology, address a longstanding challenge in stage II colorectal cancer, where the decision to offer adjuvant chemotherapy after surgery remains finely balanced, the researchers say.
While chemotherapy is routinely recommended for stage III disease, only selected stage II patients with high-risk pathological features are considered for treatment.
Even within that group, outcomes vary considerably, often leaving clinicians uncertain about which patients are most likely to benefit.
Lead author Francis Magisson, a PhD candidate at La Trobe University’s School of Computing, Engineering and Mathematical Sciences, said the AI focused on analysing the tumour’s invasive front, the leading edge where cancer cells infiltrate surrounding tissue.
“This information could be used to assist pathologists and clinicians to identify which stage II cancer patients are at higher risk of relapse and may need closer monitoring or additional treatment such as chemotherapy,” he said.
Rather than relying solely on image recognition, SÉMIL combines visual analysis with language-based descriptions of pathological features, allowing the model to recognise complex morphological patterns that resemble how pathologists interpret tissue architecture.
The researchers trained the algorithm on 1608 whole-slide images before validating it across patients treated at Austin Health, the MCO colorectal cancer cohort, and participants enrolled in the Australian-led DYNAMIC trial.
Across the three stage II cohorts, SÉMIL significantly separated patients into high- and low-risk groups for five-year relapse-free survival. Hazard ratios ranged from 2.10 to 4.73, demonstrating reproducible prognostic performance in both internal and external validation cohorts.
The model also stratified outcomes within patients already classified as high risk under current clinical guidelines. In this subgroup, hazard ratios ranged from 2.96 to 3.50 across all three cohorts, suggesting the AI could identify patients with particularly poor prognoses despite sharing conventional high-risk features with others.
The study found the strongest performance occurred when AI predictions agreed with pathologist assessment, the researchers said.
Patients classified as high risk by both methods experienced the poorest outcomes, while discordant cases had an intermediate risk of relapse, supporting the role of AI as a decision-support tool rather than a replacement for specialist pathology review.
Senior author Associate Professor David Williams, an anatomical pathologist at Austin Health and the Olivia Newton-John Cancer Research Institute, said the technology was intended to complement existing pathology workflows and not replace clinical decision-making.
“One of the biggest challenges in stage-two bowel cancer is identifying which higher-risk patients require treatment and weighing up the potential benefits of administering chemotherapy against side effects,” Associate Professor Williams said.
“Our study shows that AI-based assessment has potential to provide pathologists with an additional layer of information that could help clinicians make more informed decisions about next-stage treatment options.”
Unlike many emerging biomarkers, the model requires no additional laboratory testing. It operates using routinely prepared H&E slides, meaning it could potentially be integrated into existing digital pathology systems without new tissue sampling or expensive molecular assays.
Related
The AI also demonstrated biological validity by correctly identifying pathological features associated with known molecular subtypes of colorectal cancer, including mismatch repair deficiency, BRAF mutations, proximal tumour location and increased immune infiltration, suggesting it was capturing meaningful tumour biology rather than imaging artefacts.
Multivariable analysis showed SÉMIL remained an independent predictor of relapse after adjustment for T stage, mismatch repair status, lymphovascular invasion, lymph node assessment, and tumour-infiltrating lymphocytes, with patients classified as having infiltrative growth patterns facing almost double the risk of recurrence.
The researchers acknowledged several limitations, including the fact that the study focused on H&E-stained sections and invasive front as a single prognostic factor, and the analysis utilised a single slide per patient, which may not be representative if invasive front patterns varied across the tumour.
“While we validated across three multi-institution cohorts, all cohorts were from similar geographic regions; broader geographic and demographic validation would be needed,” they wrote.
“Furthermore, although training on the discovery cohort and validating on a different-stage population is supported by the consistent validation results, differences in tumour biology between stages could introduce bias.
“The survival stratification threshold was also selected on the discovery cohort, which carries a risk of over-optimisation.
“Finally, ground truth annotations were provided by a single expert gastrointestinal pathologist, which provides internal consistency but limits claims about reproducibility of the reference standard; multi-reader annotation studies would be needed to assess interobserver variability directly.”
However, they concluded that SÉMIL provided reliable prognostic stratification for stage II colorectal cancer using routine pathology slides, including in patients already classified as high risk under current guidelines.
They said the AI tool complemented, rather than replaced pathologist assessment by providing objective, reproducible risk estimates that captured clinically relevant tumour biology.
While the results supported its potential to improve risk assessment and treatment decision-making, prospective studies were still needed before it can be adopted in routine clinical practice, they concluded.
The study also involved research collaborators from the La Trobe Centre for Molecular Science, La Trobe School of Cancer Medicine, WEHI, Monash University, University of Melbourne, UNSW Sydney, Peter MacCallum Cancer Centre, and other leading cancer centres around Australia.
