Study Of Automatic Algorithms On Myocardial Location And Segmentation Based On Deep Learning

In recent years,cardiovascular disease has gradually evolved into a high incidence disease,which is the first cause of death from non-communicable diseases.The realization of accurate myocardial location and segmentation algorithm based on cardiac magnetic resonance images plays an important auxiliary role in the clinical diagnosis of cardiovascular diseases.Through the myocardial location and segmentation results,cardiac function evaluation indexes such as myocardial thickness and ejection volume can be measured to achieve rapid diagnosis.Due to the characteristics of low contrast?unclear borders between myocardium and surrounding tissues and irregular shape of the right myocardium in CMR images,it has brought great challenges to the accurate positioning and segmentation of myocardium.Most of the traditional methods require manual participation,and at the same time it is difficult to obtain good myocardial positioning and segmentation accuracy and robustness.Therefore,the study of automatic myocardial location and segmentation algorithm with high-precision and high-efficiency has important scientific significance and clinical application value.After analyzing the shortcomings of existing myocardial location and segmentation algorithms at home and abroad,this paper proposes two different deep learning architectures to implement myocardial location and segmentation: the myocardial location and segmentation framework based on Faster R-CNN+ISU-Net?the integrated framework of myocardial location and segmentation based on BR-Mask R-CNN.The main content and contributions of this paper are as follows:1.This paper proposes a new myocardial location and segmentation framework based on Faster R-CNN+ISU-Net.Firstly,Faster R-CNN is used to realize the myocardial location,Faster R-CNN algorithm showed good positioning performance in this task.The location accuracy m AP in the test set can reach 99.96%,and there is basically no missing or false detection.In the task of myocardial segmentation based on the U-Net algorithm,an ISU-Net algorithm that is more suitable for the CMR dataset of this paper is proposed.This algorithm reduces downsampling modules of U-Net network and adds batch normlization layer to enhance the generalization of the network and relieve the phenomenon of myocardial mis-segmentation,then use the idea of the Inception module to achieve multi-scale feature fusion at a lower cost and improve the segmentation performance of the network.The experimental results based on the Kaggle competition's myocardial data set show that the model can achieve a 73.38% DSC segmentation accuracy while greatly reducing the calculation of parameters.2.An integrated framework of myocardial location and segmentation based on BRMask R-CNN is proposed.First of all,the feature fusion path of bottom-up and horizontal short-circuit connection is added to Res Net-FPN of Mask R-CNN feature extraction network to enhance the role of the underlying features in the feature hierarchy,and the shallow detailed feature information of the image is better restored.Then,by use the idea of residual connection add a boundary refinement BR module to the mask branch of Mask R-CNN to refine the edges of the generated mask and obtain a more accurate target mask.The results based on the Kaggle myocardial test data set show that the comprehensive accuracy of location and segmentation m AP can reach 79.18%,which is 3.55% higher than the traditional Mask R-CNN algorithm.