Abstract

Background: Pancreatic ductal adenocarcinoma (PDAC) carries a dismal prognosis, and KRAS and TP53 mutational status is increasingly recognized as both prognostic and therapeutically actionable. Because next-generation sequencing has long turnaround times, high costs, and substantial tissue requirements, we aimed to develop and externally validate a deep-learning framework for inferring KRAS and TP53 mutation status directly from routine hematoxylin-and-eosin (H&E) whole-slide images (WSIs) of PDAC. Methods: A training cohort was assembled from TCGA-PAAD (n = 206) and CPTAC-PAAD (n = 147), and an independent external validation cohort (n = 86) was obtained from Osaka Metropolitan University (OMU) Hospital. We benchmarked 28 model configurations per gene, comprising three pathology foundation models (CONCH-v1.5, CTransPath, and Prov-GigaPath) crossed with nine multiple-instance-learning (MIL) aggregators (ABMIL, CLAM-SB, CLAM-MB, DSMIL, TransMIL, MeanMIL, MaxMIL, AEM, and MIL-Dropout), plus a ResNet-50 + MeanMIL baseline. Performance was evaluated by patient-level five-fold cross-validation (AUC, accuracy, precision, sensitivity, and specificity) and external AUC; attention heatmaps were generated for interpretability. Results: For KRAS, CONCH-v1.5 + MeanMIL achieved the best internal AUC of 0.717 and CONCH-v1.5 + ABMIL the best external AUC of 0.705. For TP53, CTransPath + DSMIL achieved the best internal AUC of 0.668 and CTransPath + MeanMIL the best external AUC of 0.744. Conclusions: H&E-based deep learning can infer KRAS and TP53 mutation status in PDAC with moderate but reproducible discrimination, supporting its potential as a low-cost upstream prescreening tool that triages candidates for confirmatory molecular sequencing and genotype-directed targeted therapy.

IPC Classification

Keywords