Deep Learning of Histopathology Predicts Outcomes After Surgery for Pancreatic Cancer.

Abstract

Predicting recurrence of pancreatic cancer after surgery could inform clinical decision making, including adjuvant therapies and follow-up. This study aimed to develop and validate a deep learning model using digitized whole-slide images (WSI) of histopathology. Publicly available WSI of pancreatic ductal adenocarcinoma resections from three cohorts were used for training. The model consisted of a pan-cancer foundation model to generate embeddings, mean-pooling across tissue patches, and then a fully connected neural network. Model predictions were compared with human-labeled histopathologic features and genomic alterations. The model was externally validated in a meta-analysis of a single-center cohort from Princess Margaret Cancer Centre, a multicenter cohort from France, and the PRODIGE 24 trial of adjuvant chemotherapy. The deep learning model was trained on 12,594 tissue patches from 257 patients. High-risk classifications were associated with squamous morphology, reactive stroma, tumor cellularity, and necrosis, whereas low-risk classifications were associated with tubulopapillary and conventional morphologies, as well as deserted stroma. High-risk cancers were enriched for basal-like gene expression profiles and distinct oncogenic pathways. In a meta-analysis of the external cohorts, the hazard ratio (HR) for death comparing high-versus low-risk cancers was 1.49 (95% CI, 1.25 to 1.79, <i>P</i> < .001), whereas the HR for recurrence or death was 1.41 (95% CI, 1.19 to 1.68, <i>P</i> < .001). The classifications remained prognostic among moderately differentiated cancers. An open-source deep learning model using WSI from pancreatic cancer resections generated risk classifications that correlated with histopathologic and genomic features. Classifications were externally validated in a meta-analysis of three cohorts. This model could be applied to WSI to provide individualized prognostic information for patients.