Prediction of hormone receptor status in breast cancer brain metastases using an MRI-based multimodal deep learning framework.

Introduction: Breast cancer brain metastases (BCBMs) represent a severe neurological complication affecting 10-20% of patients with metastatic breast cancer, with significant implications for treatment planning and prognosis. Accurate determination of hormone receptor (HR) status is critical for guiding personalized therapeutic strategies. However, traditional biopsy-based assessment is invasive, carries procedural risks, and may not capture the molecular heterogeneity of multiple metastatic lesions. Non-invasive methods for predicting HR status from brain MRI could transform clinical practice, yet existing approaches face three critical limitations: incomplete information utilization from single-modal features, ignored biological correlations among hormone receptors, and underexploited anatomical position information. Methods: We developed a novel multi-modal deep learning (DL) framework leveraging the BCBM-RadioGenomics dataset from The Cancer Imaging Archive (TCIA), comprising 268 MRI studies from 165 patients with expert-validated tumor segmentations, 107 PyRadiomics-derived radiomic features, and clinical HR status. Our framework introduces three key methodological components: (1) a three-modal complementary fusion architecture integrating 3D DL features extracted from tumor regions of interest using a 3D ResNet, radiomic features, and anatomical position encoding representing tumor location through hemisphere and brain region one-hot encoding; (2) a multi-task collaborative learning framework with hard parameter sharing to simultaneously predict ER, PR, and HER2 status, employing a weighted loss function to address class imbalance; and (3) a lesion-level sampling strategy to maximize data utilization while preventing data leakage. Results: The proposed model achieved area under the curve (ROC-AUC) values of 0.8763 ± 0.0393 for HER2, 0.8621 ± 0.0304 for ER, and 0.8993 ± 0.0192 for PR prediction, outperforming traditional machine learning baselines. The multi-task learning framework demonstrated improvements for the challenging PR prediction task, while the three-modal fusion architecture consistently outperformed single-modal approaches across all tasks. Discussion: This study demonstrates the feasibility and clinical potential of non-invasive molecular profiling of BCBMs using a multi-modal DL approach. These findings might offer a promising tool for treatment planning and patient management, reducing the need for invasive biopsies while enabling more accurate and comprehensive molecular characterization of brain metastases.