Advancing breast cancer relapse prediction with radiomics and neural networks: a clinically interpretable framework.

Early assessment of breast cancer relapse can significantly impact survival rates and overall oncological outcomes, highlighting the need to use sophisticated diagnostic strategies in clinical trials. This work utilizes clinically relevant radiomic features extracted from digital mammograms to develop a deep learning-based model for forecasting breast cancer relapse. Features, including tumor size, shape, margin characteristics, molecular subtype, and breast density, were systematically extracted from our private, in-house dataset, providing a comprehensive representation of intrinsic tumor properties and assisting in relapse prediction. The predictive model demonstrated outstanding performance with an average area under the curve (AUC) of 0.957, highlighting its effectiveness in identifying possible relapse. This approach not only underscores the abilities of radiomics in enhancing the granularity of tumor assessment but also assists in identifying cancer recurrence during the treatment stage, promising significant strides toward personalized cancer therapy.