TransDE: A Transformer and Double Encoder Network for Medical Image Segmentation

Abstract

Over the past decade, medical image segmentation has become a necessary prerequisite for disease diagnosis and treatment planning. The deep convolutional neural networks (CNN) have been widely adopted in medical image segmentation which achieves promising performance. However, due to the intrinsic locality of convolution operations, CNN demonstrates limitations in explicitly modeling long-range dependency. Recently proposed hybrid CNN-Transformer architectures that combine the global perception capability of local feature and the local details of global reppresentations. However, the serial structure of CNN and transformer will increase the computational complexity, and then the redundant information generated by convolution operation may leads to the failure of long-range modeling. To this end, this paper proposes a double encoder framework including global encoder and local encoder, TransDE for short, to medical image segmentation. The global encoder takes transformer that designed for sequence-to-sequence prediction, while the local encoder adopts VGG-19 combined with the atrous spatial pyramid pooling (ASPP) to bring about local feature extraction. The experimental results of enteroscopy dataset and dermoscopy dataset show the superiority of our TransDE achieving around 1.97% improvement on CVC-ClinicDB in terms of DSC and 1.6% improvement on Lesion Boundary Segmentation challenge.