Assessing the Impact of Foundation Model Features and MIL Aggregators for HER2 Status Prediction from Whole-Slide Images
Year: 2026
Type: conference-paper
Abstract: This study investigates HER2 status prediction from hematoxylin-and-eosin-stained whole-slide images using weakly supervised learning. Patch-level features extracted by two histopathology-specific foundation models, CTransPath and UNI2, were evaluated with multiple instance learning architectures across internal and external test settings. Models were trained on the Yale HER2 and HEROHE cohorts; Yale-trained models were further evaluated on the independent TCGA-BRCA cohort to assess cross-dataset generalization. Internal evaluations showed that several MIL architectures achieved competitive accuracy, whereas on the external TCGA-BRCA test set, UNI2-based models consistently preserved higher performance. In contrast, CTransPath-based models showed a marked drop. These findings suggest that, for HER2 prediction from whole-slide images, robustness of the feature representation under domain shift is more critical than increasing model complexity.
Authors: Mücahit Ertano, Abdulkerim Capar
Institution: Istanbul Technical University
Language: English
Topic: AI in cancer detection
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202162179