AI-based histopathology analysis predicts checkpoint inhibitor response in advanced melanoma and identifies patterns associated with response.

INTRODUCTION

Readily available predictive biomarkers for immune checkpoint inhibitor (ICI) response in advanced melanoma are limited. This study evaluates the predictive value of AI-based histopathology analysis.

METHODS

Patients with advanced cutaneous melanoma treated with first-line anti-PD1 ±anti-CTLA4 between 2016 and 2023 across 11 Dutch centers were retrospectively identified using prospectively collected registry data. Pre-treatment H&E-stained metastatic slides were collected from 30 pathology labs and analyzed using 14 foundation models combined with attention-based or clustering-constrained multiple-instance learning. Primary outcome was best overall response (ORR) (RECIST 1.1) assessed through leave-one-hospital-out cross-validation. Response probabilities were added to a clinical model built with AIC-based backward selection (including WHO performance status, liver metastases, LDH, ICI-type, AJCC stage, brain metastases). Feature embedding clusters were labelled by three pathologists blinded to outcomes to assess histology driving predictions.

RESULTS

Of 1935 patients, 1177 had a pre-treatment metastatic specimen available, most commonly lymph node or skin/soft tissue; 35.3% was treated with anti-PD1 +anti-CTLA4 combination therapy. ORR was 56.7% (n = 667/1177). The best AI-based model achieved an AUROC of 0.63 (95% CI:0.60-0.66) with overestimated risk. Combining the AI-based analysis with clinical information increased AUROC to 0.66 (95% CI:0.63-0.69) with improved calibration. Model predictions were driven by immune infiltration and epithelioid morphology for response, and by spindle-cell morphology, and necrosis for non-response.

CONCLUSION

AI-based analysis of diagnostic pre-treatment metastatic melanoma samples predicts ICI outcomes by using explainable histological features, such as immune infiltration, cell morphology, and necrosis. The predictive performance of the AI-based analysis improves further upon addition of clinical information.