Memory-Guided Transformer With Group Attention For Assisting Knee Injury Diagnosis Using MRI

Magnetic resonance imaging(MRI)is the preferred method for clinical examination of knee injury.However,MRI examination of knee joint requires high time and experience of doctors,so a set of intelligent auxiliary diagnosis methods is urgently needed.However,the existing intelligent assisted diagnosis methods fail to make good use of the characteristics of medical image serialization,resulting in incomplete feature information extraction,inaccurate diagnosis results,and lack of interpretable analysis.To solve the above problems,this thesis proposes an auxiliary diagnosis algorithm based on attention mechanism,and conducts interpretability analysis research at the same time,thus improving the efficiency and accuracy of knee MRI examination and reducing the misdiagnosis rate.The research content of this thesis is as follows:(1)In view of the problems of incomplete extraction of serialized feature information and low prediction accuracy by current methods,this thesis proposes Memory-guided Attention Network.The model extracts the feature information of a single sequence level and the feature information of data set level respectively by using the packet channel attention method and Memory module,and performs information fusion and result prediction based on the self-attention module.The experimental results on the MRNet dataset published by Stanford University show that the model algorithm achieves a high prediction accuracy in the diagnosis of knee injury,and the accuracy of the diagnosis of knee abnormality and ACL tear is improved by 0.24% and 1.21%,respectively,compared with the most advanced algorithms.(2)In view of the lack of interpretability analysis in current methods,this thesis proposes an interpretability analysis method based on grouping attention mechanism,gradient weighted class activation mapping and Memory module.The experimental results show that the interpretability analysis based on the grouping Attention mechanism and the Attention Heatmap generated by the network backbone module can intuitively display the basis for the network model to make diagnosis prediction,which improves the reliability and intuitiveness of auxiliary diagnosis methods(3)In view of the lack of knee image data set in China,this thesis produces and publishes the Knee MRI Data Set of University of Electronic Science and Technology of China,and verifies and compares the diagnostic effect based on the proposed model.All the case samples in the data set are from Sichuan Provincial People’s Hospital,Affiliated Hospital of University of Electronic Science and Technology of China.The data set includes 1,005 cases and includes labels for knee abnormalities,ACL tears,and meniscus tears for each case.The overall experimental results show that the algorithm maintains a high diagnostic accuracy on all data sets,and can provide correct diagnostic basis,which can better assist doctors to conduct clinical MRI examination of knee joint.