MoSID: Modality-Specific Information Disentanglement from Multi-parametric MRI for Breast Tumor Segmentation

Breast cancer is a major health issue, causing millions of deaths each year worldwide. Magnetic Resonance Imaging (MRI) is an effective tool for detecting and diagnosing breast tumors, with various MRI sequences providing comprehensive information on tumor morphology. However, existing methods for segmenting tumors from multiparametric MRI have limitations, including the lack of considering intermodality relationships and exploring task-informative modalities. To address these limitations, we propose the Modality-Specific Information Disentanglement (MoSID) framework, which extracts both intra- and inter-modality attention maps as prior knowledge to guide tumor segmentation from multi-parametric MRI. This is achieved by disentangling modality-specific information that provides complementary clues to the segmentation task and generating modality-specific attention maps in a synthesis manner. The modality-specific attention maps are further used to guide modality selection and inter-modality evaluation. Experiment results on a large breast dataset show that the MoSID achieves superior performance over other state-of-the-art multi-modality segmentation methods, and works reasonably well even with missing modalities.