Research And Implementation Of A Medical Image Segmentation Algorithm Based On Transformer With Context Aggregation

In recent years,the appearance of attention mechanism and Transformer has greatly resolved the problem of extracting global contextual information in the field of computer vision.In this context,various medical image segmentation algorithms based on Transformer have been suggested and applied.However,despite the advantage of obtaining global contextual information,there are still problems in extracting and aggregating the multi-scale contextual information from multi-scale feature maps and multi-scale segmentation targets while effectively obtaining local contextual information.To solve these problems,this paper proposes a medical image segmentation algorithm based on Transformer with context aggregation,Channel-Separate Transformer,which may solve the problems of the current segmentation algorithm based on Transformer,i.e.,the weakness in obtaining multi-scale contextual information and local contextual information.The overall structure of Channel-Separate Transformer helps to extract and aggregate multi-scale contextual information from the multiscale feature maps.Channel-Separate Self-attention modules applied in the encoder and decoder of Channel-Separate Transformer enable the extraction and aggregation of multi-scale contextual information from the multi-scale segmentation targets.Meanwhile,Local Self-attention inside each Channel-Separate Self-attention module further extracts local contextual information.Channel-Separate Transformer has been experimented in two open datasets,Synapse(a multi-organ segmentation dataset)and ACDC(Automated Cardiac Diagnosis Challenge).It is found that the current algorithm performs better that the other algorithms.Based on the above algorithm,this paper designs a medical image segmentation system based on Transformer with context aggregation,applying it to accomplish tasks of dataset selecting,parameter configuring,algorithm training,model predicting and performance evaluating.There are three layers in the system structure:interaction layer,function layer,and data layer,from top to down.The interaction layer operates the control interaction of all the graphic interfaces in the system and demonstrates graphic information.The function layer undertakes the core business logic,including training and monitoring algorithm,formulating different results of segmentation prediction,and obtaining history prediction data for evaluation.The data layer is responsible for the integration of applied data in the system,including medical image dataset and persistent data from training,predicting and evaluating.Meanwhile,through tests and analysis,this paper has verified that the needs of analyzing and demonstrating the research process and results of medical image segmentation can also be satisfied in this system.