MDE-Net: Multi-Layer Depth Extraction Network With Attention Mechanism for Medical Image Segmentation

Accurate segmentation of organ and pathological tissue images is of great significance to the diagnosis and treatment of various diseases. However, it still faces great challenges due to the inherent complexity, diversity, noise and occlusion of medical image data. To solve these problems, based on the U-Net framework, we propose a medical image segmentation algorithm named MDE-Net, which combines multi-layer deep feature extraction module and attention mechanism. Firstly, in the encoding part, we introduce the hybrid convolutional feature extraction (HCFE) module as a replacement for traditional convolutional blocks to allow for a more robust extraction of features at multiple scales and help expand the receptive field. Subsequently, we design the multi-layer pooling and channel-spatial squeeze & excitation (MPcsSE) module, which extracts more image context information by multi-layer pooling connection of the coding part and introducing csSE module in the middle connection part. Finally, in the decoder part, we design the SE-MultiResConv that combines multi-scale residual convolution with SE attention mechanism to improve segmentation accuracy and prevent the loss of detail information during up-sampling. In extensive experiments, we conducted detailed tests on two publicly available medical image datasets to rigorously evaluate the performance of our proposed MDE-Net. For the ISIC-2018 dataset, MDE-Net achieved remarkable metrics with Accuracy of 91.59%, Matthews correlation coefficient (Mcc) of 81.78%, Dice of 86.63%, and Jaccard of 76.98%. Similarly, on the COVID-19 dataset, MDE-Net exhibited outstanding performance, achieving Accuracy of 95.53%, Mcc of 79.92%, Dice of 83.43%, and Jaccard of 70.93%. The excellent performance of MDE-Net on these datasets proves its effectiveness and generalization in medical image segmentation tasks. By delivering precise and dependable segmentation outputs, MDE-Net demonstrates a transformative potential for the diagnosis and treatment of diverse medical conditions. MDE-Net’s contribution can significantly streamline diagnostic processes, minimize human error, and optimize resource allocation in clinical settings, making it a valuable tool in advancing healthcare.