DualMiT-Net: Hybrid Transformer-CNN for Breast Mass Segmentation

Researchers have introduced DualMiT-Net, a novel dual-branch deep learning architecture for breast mass segmentation in mammographic regions of interest. The model combines a local branch using a Mix Transformer (MiT-B5) encoder to capture mass shape, texture, and boundary details, with a global branch employing an EfficientNet-B5 encoder to learn surrounding breast tissue context. Features from both branches are shared at deeper encoder levels and progressively fused in a single decoder, with a spatial gate controlling the addition of global information during decoding. The study also evaluated four input representations, selecting a percentile-windowed mammogram combined with a Gabor texture response as optimal. The model was trained and evaluated on an unspecified dataset, with the paper available on arXiv under identifier 2608.15019. This work addresses challenges in computer-aided mammography, such as low contrast, irregular shapes, and ambiguous boundaries of breast masses, aiming to improve segmentation accuracy for clinical applications.