A 3D-2D Hybrid Network with Regional Awareness and Global Fusion for Brain Tumor Segmentation

Accurate segmentation of brain tumors in MR images is crucial for their clinical diagnosis and treatment. Some existing methods do not adequately consider the relationship between tumor regions and the effect of fuzzy boundaries on segmentation. In this paper, we propose a 3D-2D Hybrid Network with Regional Awareness and Global Fusion for Brain Tumor Segmentation (HRGBTS). The model consists of three components: a hybrid encoder, a regional awareness module, and a feature fusion decoder. Specifically, the hybrid encoder uses 3D-2D hybrid convolutional blocks to extract multi-scale features from the brain tumor region. The regional awareness module (RAM) enhances the segmentation of tumor sub-regions by dynamically sensing the relationship between tumor cells and surrounding tissue cells through graph convolutional interactive inference. In the decoding stage, the feature fusion decoder, which consists of global fusion modules (GFMs) and cross-dimensional skip connections, is designed to improve boundary reconstruction. The GFM effectively fuses tumor information, enabling better tumor boundary reconstruction and addressing the challenge of boundary blurring. The cross-dimensional skip connections address the information mismatch problem caused by cross-dimensional changes. Extensive evaluations were made on three benchmark datasets, BraTS2018, BraTS2020, and BraTS2021, the Dice coefficients in the whole-tumor region (WT) were achieved to be 0.906, 0.917, and 0.903, respectively.