Three-branch dynamic selection network for 3d medical image segmentation

Abstract 3D medical images present significant challenges for segmentation due to the diversity of anatomical structures, as well as regions with low contrast and noise. To address these issues and achieve precise segmentation, we propose a Three Branch Dynamic Selection Network (TB-Net). This network enhances segmentation performance by capturing both rich detailed features and abstract semantic features. TB-Net consists of three parallel feature extraction paths: a CNN branch for capturing local texture and spatial features, a Transformer branch for modeling global semantic information, and a Hidden Connection (HC) branch for learning and reusing historical information from the encoding stage. To efficiently integrate the extracted features, we introduce a Dynamic Selection (DS) module that adaptively fuses multi-branch features. We conducted quantitative experiments on four MRI and CT datasets (OAI-ZIB, Hippocampus, Parse, and Spleen), and TB-Net demonstrated overall superior performance compared with the baseline methods across these datasets. On the OAI-ZIB dataset, TB-Net achieved an average segmentation performance of 90.2% ± 0.02 Dice, 0.5 ± 0.1 mm ASSD, and 1.7 ± 0.4 mm HD95 for femoral cartilage, representing improvements of 0.9% in Dice, a 0.1 mm decrease in ASSD, and a 0.2 mm decrease in HD95 compared with the second-best method.