An Improved Segmentation Network Based on an Encoder-Decoder Architecture

U-shaped encoder-decoder convolutional neural networks have shown significant success in medical image segmentation. However, these networks often face challenges, including detail loss in skip connections, insufficient integration of multi-scale spatial information, and limited feature extraction in the encoder. To address these limitations, this study introduces an improved segmentation network for brain tumor segmentation tasks. The proposed model incorporates a Dual Cross-Attention (DCA) mechanism into the skip connections, enhancing the representation of original image details by supplementing initial feature information to the decoder. Additionally, a redesigned Depthwise Separable Convolution Module (DSC-M) replaces conventional convolutional layers in the encoder and decoder, reducing model parameters while maintaining performance. The network was validated on the BraTS2021 brain tumor dataset, achieving performance metrics surpassing state-of-the-art models.