Brain Tumor Segmentation Algorithm Based on Asymmetric Encoder and Multimodal Cross-Collaboration

Abstract

To address the challenges of insufficient multimodal information fusion and insufficient long-range dependencies features extraction for brain tumor segmentation, this paper propose a novel network based on asymmetric encoder and multimodal cross-collaboration. The network employs an asymmetric encoder-decoder architecture. Firstly, the invert ConvNext split convolution (ICSC) block is used in the local refinement encoder and improved SwinTransformer with DscMLP enhancements (DscSwinTransformer) module is used in global associative encoder. The local and long-range dependencies of each stage of two parallel encoders can be well extracted by hybrid fusion. Moreover, this paper adds a multimodal cross-collaboration (MCC) module at the beginning of the two encoders to fully exploit the complementary information between modalities and reduce the reliance on a single modality during model training. Coordinate Attention (CA) is used in the bridge part of the encoder and decoder to capture important spatial location information. Then, the depthwise separable convolution (DscConv) module is used in the decoder branch to reduce the computation while maintaining good feature extraction ability. Finally, this paper uses a hybrid loss function of BCE, Dice and L2 loss to mitigate the problem of class datas imbalance. Experimental results show that our model achieves Dice coefficients of 0.897, 0.905 and 0.824 in the whole, core and enhanced tumor regions, respectively. These results show that the performance of our proposed method outperforms in comparison with several existing methods in core and enhanced tumor regions.