Medical Image Segmentation Based on Transformer and HarDNet Structures

Abstract

Medical image segmentation is a crucial way to assist doctors in the accurate diagnosis of diseases. However, the accuracy of medical image segmentation needs further improvement due to the problems of many noisy medical images and the high similarity between background and target regions. The current mainstream image segmentation networks, such as TransUnet, have achieved accurate image segmentation. Still, the encoders of such segmentation networks do not consider the local connection between adjacent chunks and lack the interaction of inter-channel information during the upsampling of the decoder. To address the above problems, this paper proposed a dual-encoder image segmentation network, including HarDNet68 and Transformer branch, which can extract the local features and global feature information of the input image, allowing the segmentation network to learn more image information, thus improving the effectiveness and accuracy of medical segmentation. In this paper, to realize the fusion of image feature information of different dimensions in two stages of encoding and decoding, we propose a feature adaptation fusion module to fuse the channel information of multi-level features and realize the information interaction between channels, and then improve the segmentation network accuracy. The experimental results on CVC-ClinicDB, ETIS-Larib, and COVID-19 CT datasets show that the proposed model performs better in four evaluation metrics, Dice, Iou, Prec, and Sens, and achieves better segmentation results in both internal filling and edge prediction of medical images. Accurate medical image segmentation can assist doctors in making a critical diagnosis of cancerous regions in advance, ensure cancer patients receive timely targeted treatment, and improve their survival quality.