Attention-based multiple-instance learning enables whole-slide pathology classification of esophageal squamous cell carcinoma versus esophageal adenocarcinoma with quantitative explainability

Background Accurate histological subtyping of esophageal cancer is critical for treatment selection and prognostication. Whole-slide images (WSIs) provide rich morphological information, but they are extremely large and lack region-level annotations. We evaluated an interpretable deep learning pipeline for classifying esophageal squamous cell carcinoma (ESCC) and esophageal adenocarcinoma (EAC) using WSIs. Methods This retrospective study included 139 hematoxylin-and-eosin-stained WSIs from 139 patients (58 with EAC and 81 with ESCC). To differentiate EAC from ESCC, tissue tiles were generated and fixed-length embeddings were extracted using the original frozen Google ViT. During model fitting, each slide was represented by a randomly sampled bag of 128 tile embeddings and slide-level prediction was performed using an attention-gated multiple-instance learning model. We employed focal loss and weighted sampling strategies and evaluated internal performance using stratified fivefold cross-validation. The five original MIL checkpoints were applied to a balanced 50-case non-TCGA public cohort with an internal-only Platt mapping and post hoc full-cohort threshold optimization. Results The mean AUC was high in the training sets (0.98 ± 0.01) and internal validation folds (0.95 ± 0.05), with the best validation fold exhibiting an AUC of 0.99 and an accuracy of 96.4%. Neither normalized attention entropy ( p = 0.364) nor maximum attention ( p = 0.423) differed significantly between histologies. In post hoc full-cohort threshold optimization, internally calibrated external scores achieved apparent 80.0% accuracy and balanced accuracy, with adenocarcinoma sensitivity of 76.0% and ESCC sensitivity of 84.0% . Conclusions This model demonstrated potential as an interpretable slide-level screening approach for ESCC versus EAC classification, but the external operating threshold was domain dependent. Prospective multicenter validation using a prespecified cutoff is required.