Medical Image Segmentation Based On Group Transformer

The retinal vessels in fundus images play an important role in the screening,diagnosis and treatment of a wide range of diseases.The complex morphology of the retinal vessels,the number of capillary branches and the presence of lesions that interfere with them make manual segmentation of the vessels time-consuming and difficult and make accurate segmentation and judgment difficult.However,the computer-aided diagnosis system can achieve accurate segmentation of retinal vessels,reduce the manual burden and improve the efficiency of diagnosis,which has good application value and prospects.To this end,this paper addresses the above-mentioned issues,and the specific work is as follows:(1)A Transformer and CNN-based retinal vessel segmentation method is proposed for the problems of uneven illumination,high noise and complex structure in fundus images.Firstly,the fundus image is pre-processed to enhance the contrast of blood vessels,attenuate the effect of noise in the fundus image,and expand the sample size.Secondly,the reuse of features is achieved by replacing the normal convolution with an aggregated residual module using U-Net,a classical framework in medical image segmentation.Then,an attention mechanism is added to the jump connections of the network,which is used to capture more features and preserve more spatial information.Experimental results show that the method has an accuracy of 96.27% and 96.69% in the DRIVE and STARE datasets,improving image quality and enabling better vessel segmentation performance.(2)A multi-scale Transformer-based retinal vessel segmentation method is proposed in order to segment retinal vessels more accurately and to improve the segmentation of fine vessels.Firstly,multi-scale images are used as input,and the residual blocks are used to convey feature information more effectively.Deformable convolution and multi-head self-attention mechanisms are used for the structural properties of the vessels to perform effective feature refinement and achieve more accurate segmentation.Next,a new decoder module is constructed in the network to better retain feature information.Finally,a side output layer is used to complete the final segmentation.Experimental results show that the method achieves 96.31%,97.03%and 97.37% accuracy in the DRIVE,STARE and CHASE＿DB1 datasets,improving segmentation accuracy and segmenting more capillaries.(3)A high-resolution network-based retinal vessel segmentation method was proposed to address the problems of blurred vessel boundaries and discontinuous segmentation of fine vessels in fundus images.The method is based on a modified architecture of HRNet,which can maintain high resolution and has the feature of multiple information complementation for high and low-resolution representations,enhancing the representation capability of the network.It also introduces techniques such as multi-scale space,null convolution and attention mechanism into the model to extract multi-scale features and deep information from the image and improve the accuracy of segmentation of tiny blood vessels.The experimental results show that the method has 96.95%,97.36% and 97.43% accuracy in DRIVE,STARE and CHASE＿DB1 datasets,and has better connectivity for segmentation of fine vessels.