Artificial intelligence–guided microsatellite instability classification in colorectal cancer (CRC): EfficientNet-based histopathology analysis for precision treatment planning.

Abstract

e15621 Background: Microsatellite instability (MSI) represents a biomarker in CRC, affecting 15–20% of cases and guiding treatment/prognosis. MSI-high (MSI-H) tumors exhibit defective mismatch repair showing hypermutation, immunogenicity, and strong responses to ICIs (pembrolizumab, nivolumab) but poor traditional chemotherapy outcomes. Conversely, microsatellite stable (MSS) tumors respond better to fluorouracil-based regimens. Current MSI testing via immunohistochemistry or PCR is labor-intensive, time-consuming and variable. Automated deep learning systems capable of predicting MSI status directly from routine histopathology slides could accelerate diagnostics, optimize therapy, and reduce treatment delays.OBJECTIVES:To develop and validate a deep learning framework for predicting MSI status from CRC histopathology; to evaluate lightweight model performance via knowledge distillation from a higher-capacity teacher; and to assess global feasibility using expert-reviewed, multi-institutional data. Methods: We assembled 10,000 anonymized hematoxylin-eosin stained histopathological images of colorectal tissue from multi-institutional sources. MSI status was confirmed by gold-standard PCR or immunohistochemistry; pathologist confirmed ground truth. After preprocessing and stain augmentation, images were stratified into training (60%), validation (20%), and testing (20%) cohorts. EfficientNetB4 (17.4M parameters, 380×380 resolution), selected for compound resolution scaling, was trained as reference. EfficientNetB1 (7.8M parameters, 240×240 resolution, 55% parameter reduction) underwent structured knowledge distillation using soft targets from EfficientNetB4. Model performance was evaluated on accuracy, sensitivity, specificity, F1-score, and AUROC for MSI classification. The optimized EfficientNetB1 was deployed in a digital pathology platform and independently assessed by 47 pathologists across 6 continents. Results: EfficientNetB4 achieved high diagnostic accuracy for MSI status prediction. Distilled EfficientNetB1 demonstrated comparable performance, with accuracy exceeding 93% on test data, AUROC > 0.91, balanced sensitivity (94.2% MSI-H, 92.1% MSS), and specificity > 93%. Performance remained stable across geographically distinct datasets with staining protocol variation. Pathologist evaluators reported the system as valuable for MSI triage and treatment planning support. Conclusions: EfficientNet enables accurate MSI prediction from histopathology. Knowledge distillation from EfficientNetB4 to EfficientNetB1 preserved diagnostic accuracy while cutting parameters 55%, supporting rapid stratification for immunotherapy. Prospective evaluation in workflows is warranted to accelerate CRC precision oncology.