HSegFormer: hybrid CNN–transformer with stage attention for brain tumor MRI segmentation

Abstract

Brain tumor segmentation from contrast-enhanced T1-weighted MRI remains challenging because of heterogeneous tumor appearance and complex boundaries. This study proposes HSegFormer, a hybrid CNN–Transformer architecture that combines convolutional feature extraction with transformer-based contextual modeling, attention-guided decoding, and deep supervision. The model is trained using a combination of binary cross-entropy and Dice losses and evaluated on the BRISC 2025 and Figshare datasets. Experimental results show that transformer-based and hybrid architectures achieve higher IoU and Dice scores than the evaluated CNN-based models. Among the evaluated methods, HSegFormer achieves the highest segmentation performance while maintaining competitive computational efficiency. Ablation studies indicate that stage-wise attention, deep supervision, and the proposed Combo Loss each contribute to the overall performance. These results suggest that integrating convolutional and transformer representations can improve brain tumor segmentation across the evaluated datasets.