Prediction of KRAS mutation status from H&E foundation model embeddings in non-small cell lung cancer

Marc Robbins, Jessica Loo, Saurabh Vyawahare, Yang Wang, Carson Mcneil, Sudha Rao, Pok Fai Wong, Ehud Rivlin, Shamira Weaver, Roman Goldenberg

Keywords: KRAS, NSCLC, multiple instance learning, foundation model, adenocarcinoma, squamous cell carcinoma, histopathology, mutation

TL;DR: We predicted KRAS mutation status in NSCLC from foundation model embeddings of H&E slides using a variety of methods

Abstract: We predicted KRAS mutation status on non-small cell lung cancer (NSCLC) H\&E images from foundation model embeddings. We evaluated a variety of attention-based multiple instance learning (MIL) models and aggregation strategies for a tilewise linear classifier. MIL with self-attention performed the best (AUC=0.822) followed by the minimum over tiles classified with the linear model (AUC=0.810). Self-attention was necessary for MIL to surpass tilewise linear classification when a wide range of aggregation techniques was considered.

Submission Number: 16