MulFF-Net: A Domain-Aware Multiscale Feature Fusion Network for Breast Ultrasound Image Segmentation With Radiomic Applications

Ultrasound is a widely used, nonionizing, and cost-effective imaging modality for breast cancer screening. However, its inherent limitations, such as low contrast, variable lesion sizes, and indistinct lesion boundaries, pose significant challenges for precise segmentation. To mitigate these issues, we proposed a domain-aware multiscale feature fusion network (MulFF-Net), a novel approach for breast ultrasound (BUS) image segmentation. The fundamental novelty of MulFF-Net lies in the incorporation of a dilated convolution spatial attention module (DCSAM) into the skip connections. Unlike conventional skip connections, this advanced design captures multiscale contextual information with spatial attention and enhances the model’s ability to accurately localize segmentation regions. Furthermore, MulFF-Net uses a domain-specific encoder coupled with both channel and spatial attention mechanisms, further refining the feature representation to improve segmentation performance. MulFF-Net was evaluated on three BUS datasets, achieving superior intersection over union (IoU) and Dice scores compared with state-of-the-art models. In addition, the generated lesion masks demonstrated potential utility in downstream radiomic analysis, offering a valuable tool for further clinical applications.