Research On 3D Left Atrium Segmentation Algorithm Based On Semi-supervised Learning

Atrial fibrillation caused by ectopic beats in the left atrium can lead to complications such as vascular blockage,cardiac dysfunction,and sudden cardiac death.Therefore,accurately segmenting the left atrium from MRI images is a key step in evaluating left atrial function and diagnosing atrial fibrillation in the future.Due to its ability to utilize limited annotated data and a large quantity of unannotated data for learning,the semi-supervised learning-based left atrium segmentation method has become one of the research focuses.At present,most image segmentation algorithms can only process two-dimensional images,so it is necessary to first make two-dimensional slices along a certain plane of the threedimensional left atrial MRI image;Although some methods can directly process 3D images,due to the limited number of left atrial MRI images,the complete MRI image needs to be cropped into 3D sub images for training,resulting in segmentation algorithms being unable to extract the overall structural information.To solve these problems,this paper studies the image segmentation algorithm of semi supervised learning,aiming to accurately segment the left atrium from 3D MRI images.The specific research work is as follows:(1)A semi supervised image segmentation algorithm of left atrium based on transfer learning is proposed.Based on the main idea of transfer learning,this algorithm makes full use of the similarity between 3D medical sub images and 3D medical images to solve the problem of small number of 3D medical images and the contradiction that the model cannot learn a complete 3D medical image.In addition,a mixed hole convolution module was designed to improve the performance of V-Net.By calculating the consistency between the segmentation task and the regression task,dual task consistency is achieved.The experimental results show that through transfer learning and mixed hole convolution,the algorithm’s indicators based on spatial overlap and spatial distance have been effectively improved.(2)A semi-supervised image segmentation algorithm for left atrium based on mean teacher model is designed and implemented.The network backbone of this algorithm is DFV-Net,which is an improvement on V-Net.DFV-Net not only deepens the network by introducing lightweight measures such as depthwise separable convolutions,but also replaces the original horizontal connections with full-scale skip connections and adds a deep supervision module.In addition,this algorithm introduces the average teacher model,which uses the exponential moving average of the academic model parameters to obtain a teacher model for prediction,and combines it with dual-task consistency to increase perturbations at both the network model level and task level.Experimental results show that by using DFVNet and the average teacher model,the final algorithm achieved improvements based on spatial overlap metrics.Both algorithms were tested on the 2018 Left Atrial Segmentation Challenge dataset and compared with state-of-the-art algorithms.The results show that the two algorithms have good segmentation performance,and prove their effectiveness and progressiveness.The former pays more attention to the boundaries of segmentation,while the latter pays more attention to the overlapping parts of segmentation.This study not only improves the accuracy of left atrial image segmentation,but also provides new ideas and successful case studies for other medical image segmentation.