GFC-Unet: An optimized 3D U-Net-based method for brain tumor segmentation

Accurate brain tumor segmentation is critical for diagnosis and treatment planning. However, U-Net-based models struggle with complex tumor morphologies and fine boundaries. This paper proposes a novel brain tumor segmentation network, termed GFC-Unet. Built upon the 3D U-Net architecture, this network introduces three core modules: (1) the Multi-modal Feature Recalibration (MFR) module; (2) the Gated Feature Convolution (GFC) module; and (3) the Feature Uncertainty Evaluation (FUE) module. Experimental results on the Brain Tumor Segmentation (BraTS) 2020 and 2021 datasets demonstrate that GFC-Unet achieves Dice coefficients of 92.82%, 87.74%, and 87.47% on BraTS2020, and 92.61%, 87.49%, and 86.89% on BraTS2021 for whole tumor (WT), tumor core (TC), and enhancing tumor (ET), respectively. Ablation studies validate the effectiveness of each module. Furthermore, extended experiments on the Data Science Bowl (DSB) 2018 cell nucleus segmentation and International Skin Imaging Collaboration (ISIC) 2018 skin lesion segmentation tasks demonstrate the strong generalization capability. By integrating these modules, GFC-Unet improves segmentation accuracy and boundary delineation, demonstrating promise for supporting clinical decision-making.