Comparative Study of Radiomics vs. Pathomics for Breast Cancer Subtype Prediction

Abstract

Breast cancer remains a leading cause of cancer-related morbidity and mortality among women worldwide, necessitating more accurate and non-invasive methods for early diagnosis and molecular subtype classification. This study presents an integrative framework combining radiomics and pathomics to enhance the precision of breast cancer subtype prediction. Radiomic features were extracted from medical imaging modalities, capturing tumor heterogeneity, shape, and texture, while pathomic features were derived from digitized histopathology slides, quantifying cellular morphology and tissue architecture. A series of machine learning models, including Random Forest, SVM, and XGBoost, were applied to unimodal datasets, with XGBoost achieving the highest accuracy of 84% (AUC: 0.88) using radiomic features and 82% (AUC: 0.85) with pathomic features. Multimodal fusion models outperformed unimodal approaches, with the Fusion-CNN model reaching an accuracy of 88% and an AUC of 0.91. SHAP analysis revealed that Texture_Homogeneity, Mitotic Count, and GLCM_Correlation were the most influential predictors across modalities. Visualizations further demonstrated clear separability between subtypes such as Luminal A and Triple Negative, validating the discriminative power of the selected features. The integration of radiomics and pathomics not only improved classification performance but also offered greater biological interpretability and clinical relevance. This study highlights the transformative role of AI-driven, multimodal diagnostic pipelines in supporting precision oncology and individualized patient management in breast cancer care.