MTF-Net: a brain tumor segmentation network with multi-scale transformer and feature fusion enhancement

To address the issues of insufficient multimodal information fusion and inadequate long-range feature extraction in brain tumor segmentation, we propose a brain tumor segmentation network based on multi-scale Transformer and feature fusion enhancement (MTF-Net). In the second to fourth encoding layers of MTF-Net, the multi-scale and spatially-aware Transformer module (MSFormer) effectively integrates multi-scale deep convolution with spatially-aware MLP to perform local-global joint modeling under different receptive fields. Additionally, the last layer of the encoder employs a Channel-Enhanced Bi-level Attention Module (CEBAM), which combines a dual-layer sparse attention mechanism with a channel enhancement structure to enhance the capability of expressing long-range dependencies. Then, in the bridging section of the encoder-decoder, a Cross-Modal Feature Interaction Module (CFIM) is used to guide the fusion of features from MSFormer and the convolutional branch, improving boundary perception and consistency across modalities. Finally, we uses a hybrid loss function composed of binary cross entropy (BCE) and Dice Loss to alleviate the issue of class imbalance. Experimental results on the BraTS2019 brain tumor dataset show that the model achieves average Dice coefficients of 0.926, 0.921, and 0.896 for whole tumor, tumor core, and enhancing tumor, respectively. Experimental results on the BraTS2021 brain tumor dataset show that the model achieves average Dice coefficients of 0.918, 0.923, and 0.901 for whole tumor, tumor core, and enhancing tumor, respectively. Compared with the evaluated segmentation methods, the proposed model achieves competitive segmentation performance across the three tumor subregions.