Research On Segmentation Of Left Atrium Based On Deep Learning

Atrial fibrillation(AF)causes blood clots to clog blood vessels,increases the risk of cardiac failure and stroke,and makes morbidity and mortality higher.The lack of a deeper understanding of atrial anatomy is one of the important reasons for the poor clinical treatment of atrial fibrillation.Although the atrium can be reconstructed by manual segmentation to explore its structure,it requires specialized domain knowledge and a large amount of labor costs.Therefore,it is of great significance to use the intelligent segmentation of the left atrium in the image to assist the doctors in the treatment of atrial fibrillation.The main research content of this project is to realize automatic segmentation of the left atrium in MRI images based on deep learning,which is divided into two parts.In the first part,an embedded model with double U structure is studied.Firstly,the up-sampling and down-sampling modules,which both look like a U shape,are built respectively as encoders and decoders of the model,and then connect these modules to form multiple paths from input to output to increase the information transmission capacity of the model.The performance comparison on the test set shows that this algorithm can effectively improve the segmentation accuracy of the model.In the second part,we introduce the attention mechanism into the model with an encoderdecoder structure,use the global average pooling of location attention to improve the ability of capturing context features and the channel attention to strengthen or weaken the choice of feature channels,then we combine them to obtain mixed features.Finally,a boundary segmentation network is introduced to improve the ability of the model to classify the boundary pixels of the left atrium.The multi-path structure introduces two types of weighted average loss functions to improve the training effect.Experiments show that these algorithms,which based on U-Net model,successively composited with attention mechanism,pre-training of encoder and boundary segmentation network,can make the performance of the model improve gradually.The results of this research show that UU-Net model and mixed feature network proposed in this paper can get the highest scores 0.8764 and 0.8841 for Dice index respectively on the dataset of magnetic resonance image of the left atrium.The comparison between the predicted result and artificial segmentation of 3D reconstruction for left atrium shows both models can meet the basic requirements for left atrium segmentation,which has a certain reference significance for doctors to analyze left atrial images and make the scheme for treating atrial fibrillation.