Key Technology Research Of Computer-aided Diagnosis For Atrial Fibrillation Based On Magnetic Resonance Imaging

Atrial fibrillation can reduce the pumping function of the heart,leading to thrombosis,heart failure,myocardial infarction,and other complications,which seriously harm people’s health.Studies have shown that changes such as left atrial enlargement,muscle hypertrophy,and fibrosis are closely related to the occurrence of atrial fibrillation.In clinical practice,manual segmentation of the left atrium from medical images is a time-consuming and laborious task,which also requires doctors to have a high degree of professional knowledge and experience.Therefore,studying the automatic segmentation of the left atrium by computer-aided diagnosis technology is of great significance for the diagnosis and treatment of atrial fibrillation.Compared with the traditional method of left atrium segmentation,deep learning models can achieve more accurate and efficient segmentation.Left atrium segmentation based on deep learning has become a research hotspot and has made remarkable progress.However,in medical images,the left atrium is closely connected with the surrounding structure and is highly similar in organization,making it difficult to distinguish the boundary and easy to mislead segmentation.The segmentation results of existing algorithms have rough edge details.In addition,obtaining and labeling medical images is difficult,resulting in a large number of unlabeled data or mislabeled data,which are called "noise data".To solve these problems,this paper proposed two left atrial segmentation algorithms,and the main research work is as follows:(1)Based on the U-Net network,a segmentation model combining residual error and attention mechanism is proposed.First,during down-sampling and up-sampling operations,the input of the previous layer and the output after two convolutions are added as the input of the next layer.This residual connection mode can gradually extract higher-level features while retaining lower-level features,which solves the problem of performance degradation when the network deepens.Then,the attention gate module is added to the tail of each jump connection,so that the model can focus on the feature learning of the segmentation target,better distinguish the boundary,and improve the accuracy of segmentation.Experiments were conducted on the 2018 left atrial Segmentation Challenge dataset,and the dice similarity coefficient,jaccard coefficient,and hausdorff distance three evaluation indexes reached 91.56%,84.50%,and 9.12,respectively,which were superior to other segmentation algorithms used for comparison,proving the excellent segmentation performance of this model.(2)Based on the MTCL network,a new semi-supervised segmentation model is proposed by taking the simplified V-Net network as its backbone network.To solve the problem that the original MTCL network is not effective in processing 3D medical images,the backbone network of MTCL is replaced by V-Net instead of U-Net.However,the complexity of the V-Net network structure and the number of parameters will result in large computation,so the number of convolutional layers in the three layers of the network is reduced,and the Dropout operation is added to simplify the V-Net network structure.The experiment was conducted on the 2018 left atrial Segmentation Challenge dataset.When the proportion of clean data in the training set was 10%,the dice similarity coefficient,jaccard coefficient,and hausdorff distance reached 88.02%,78.72%,and 10.89,respectively,which was better than other semi-supervised methods used for comparison.It is proved that it has good segmentation ability when learning with noise.