Foundation Model-Based Computational Pathology Predicts Breast Cancer Biomarker Status from Intraoperative Frozen Sections.

Prediction of breast cancer biomarkers from routine formalin-fixed, paraffin-embedded (FFPE) hematoxylin and eosin (H&E) images has been explored previously, but whether biomarker-associated morphologic signals can be reliably captured from artifact-prone intraoperative frozen-section (FS) specimens remains insufficiently studied. This retrospective proof-of-concept study developed a foundation model-based multiple-instance learning framework to predict estrogen receptor (ER), progesterone receptor (PR), human epidermal growth factor receptor 2 (HER2), and Ki-67 status directly from H&E-stained intraoperative FS whole slide images of 248 patients with breast cancer, using postoperative immunohistochemistry and fluorescence in situ hybridization as ground truth. Exploratory feature evaluation identified UNI2-h as a suitable encoder for downstream modeling. In fivefold cross-validation, areas under the receiver operating characteristic curve were 0.847 for ER, 0.906 for PR, 0.893 for HER2, and 0.799 for Ki-67, with moderate agreement for ER, PR, and HER2 and fair agreement for Ki-67. An exploratory within-cohort cross-preparation analysis showed limited FFPE-to-FS transferability, whereas the FS-trained model retained relatively favorable performance on FFPE images. Attention heat maps highlighted tumor regions with biologically plausible biomarker-associated morphologic patterns. These findings provide proof of concept that intraoperative FS H&E images retain learnable biomarker-associated morphologic signals despite preparation-related alterations.