RTNet: a residual t-shaped network for medical image segmentation

Accurate segmentation of lesion areas in medical image analysis can assist clinicians develop more personalized treatment tools, improving treatment efficacy and survival rates. Furthermore, the single encoding structure within the U-shaped architecture limits the network’s capacity to aggregate semantic information of various scales. In this paper, we propose a residual T-shaped network (RTNet) for medical image segmentation. The RTNet incorporates Multi-Level Feature Enhanced Residual Block (MFER Block) and attention gates into a T-shaped structure, making our network capture richer contextual information at multiple scales. First, by utilizing two sets of encoders with different structures, our network can extract image features at different scales. Secondly, the MFER Block is employed to combine and learn the feature representations from the encoders, facilitating more effective feature fusion. Finally, the attention gates fuse those feature maps obtained from each stage via skip connections, then they are concatenated by the decoder to generate the final feature map for segmentation. Experimental results show that, on the BUSI and ISIC2017 datasets, the IoU of our RTNet reached 67.39% and 84.39%, respectively; and the DSC score achieved 79.95% and 91.44%, respectively. Our network outperforms state-of-the-art models for medical image segmentation.