Microsatellite instability (MSI) decides how colorectal cancer is treated. MSI tumours often respond well to immunotherapy, while microsatellite stable (MSS) tumours are usually treated with conventional chemotherapy. Status is established in the laboratory by PCR or immunohistochemistry, which costs time and capacity on every case that comes through.
Reading the status off the slide
The model predicts MSI status directly from H&E-stained whole-slide images using attention-based multiple instance learning, so it can be trained on slide-level labels alone. It was trained on 1,242 cases from three public archives and evaluated on 1,094 cases from five independent external cohorts.
Three choices separate it from the usual design: the average of the foundation model’s register tokens is used as input rather than the classification token, only the highest-attention patches are kept, and large ensembles of models are combined into a single prediction.
Knowing when not to answer
Every prediction carries an ensemble confidence interval, and cases whose interval crosses the decision threshold are rejected rather than guessed. That turns the model into a triage step: high-confidence MSS cases are cleared, and everything else is forwarded to conventional testing.
Across the five external cohorts it reached sensitivities above 97% while covering 59% to 78% of cases, with negative predictive values between 98% and 100%. It reaches that on roughly a tenth of the training data used by the largest comparable study.
Where it fits
The clinically appropriate use is safety-oriented triage of MSS cases rather than full pre-screening. Accepting high-confidence MSI predictions as well raised coverage by only about 1% and introduced false positives, which is hard to justify when an MSS case labelled MSI changes treatment. The same approach extends to other tumour types and to other molecular targets.