Multi-Branch Decoding Medical Image Segmentation Based on Transformer

Abstract

Highly accurate medical image segmentation model is crucial to assist doctors in medical diagnosis and treatment. Despite the good performance of U-Net series networks in most medical segmentation tasks, they face several limitations, such as module fails to extract sufficient features and lacks ability to capture contextual information. Therefore, this paper proposes a multi-branch decoding model. This model uses Res2Net as backbone and captures features with different receptive fields through a multi-scale context-awareness module. Furthermore, during the decoding process, a Transformer decoding branch is incorporated to perceive global information. And a high-level feature aggregation decoding branch is introduced to reduce the influence of background noise. The proposed model is experimented on CVC-ClinicDB, Kvasir-SEG and ISIC 2016. Results show that the improved model outperforms the original U-Net series networks in terms of Dice and IoU.