DPF-Unet: a CNN-swin transformer fusion network for 3D brain tumor segmentation in MRI images
摘要
Three-dimensional (3D) brain tumor segmentation is a crucial task in medical imaging analysis, with significant implications for precision medicine, disease monitoring, and pathology research. However, the complex structural heterogeneity and scale variations of brain tumors in MRI data pose challenges for accurate 3D segmentation. To address these limitations, we propose DPF-Unet, a dual-path U-shaped network designed to synergize the complementary strengths of convolutional neural networks (CNNs) and Transformers for robust tumor segmentation. The CNN path captures fine-grained local details, while the Swin Transformer path models long-range global dependencies, enabling comprehensive feature learning. To unify these dual-path features, we introduce an attentional feature aggregation module that dynamically fuses channel-wise and spatial information across local and global contexts. Furthermore, we develop a multi-level feature fusion module to enrich hierarchical encoder outputs, enhancing segmentation robustness to extreme tumor scale variations. We conducted comprehensive performance evaluations and cross-domain generalization tests on DPF-Unet on the BraTS 2020, BraTS 2021, and BraTS-Africa datasets. The experimental results show that the model exhibits competitive performance on all three datasets, especially on the BraTS 2021 validation set, with Dice Scores of the ET, TC, and WT reaching 84.88