Research On Left Atrial Image Segmentation Method Based On Semi Supervised Deep Learning

Medical imaging technology has made significant contributions to the progress of modern medical research.Thanks to the rapid development of imaging technology,doctors can intuitively see the location of diseases in patients’ bodies and diagnose their condition based on the observed image content.When doctors perform diagnostic work based on medical images,they often subjectively segment medical images into meaningful parts,which is both time-consuming and laborious,and the segmentation results vary from person to person.The use of semantic segmentation technology for automatic and accurate segmentation of medical images to assist doctors in diagnosis not only greatly improves doctors’ work efficiency,but also reduces the probability of misjudgment.Semantic segmentation divides raw medical image data into structured and meaningful regions,enabling further image analysis and quantification.This technology has promoted the application and development of medical imaging,including X-ray,B-ultrasound,CT,and MRI.The number of original medical images is huge,but meaningful annotation data is very scarce,making semi supervised deep learning methods a popular research direction.In order to solve the above problems,this article mainly studies a semi supervised left atrial image segmentation method.In the current field of semi supervised deep learning,the combination of CNN and Transformer for cross learning has been proven to be a simple and effective method.The combination of Unet and Swin Net using Cross Pseudo Supervision(CPS)strategy for cross learning outperforms the original CPS network.This proves that CNN and Transformer can compensate each other during the cross learning process.Based on this background,the main research work of this article is as follows:(1)Heterogeneous cross pseudo supervised networkThis article combines Unet and Swin Unet for cross learning according to the CPS strategy.In order to alleviate the poor performance of Swin Unet when the data volume is small,the Shifted Patch Tokenization module is embedded in Swin Unet to enrich the image space information and indirectly improve the Receptive field of Swin Unet.Due to the fact that this network model contains two networks with completely different architectures,this article refers to the network as a heterogeneous cross pseudo supervised network.(2)A heterogeneous cross pseudo supervised network based on Fourier transform network.Construct a Fourier transform based auxiliary network in the input part of a heterogeneous cross pseudo supervised network to generate an auxiliary image for each input image.Simultaneously feed the original and auxiliary images into the network training and generate pseudo labels.Calculate the KL divergence of the network for predicting auxiliary and original images,used to measure the quality of pseudo labels.The network calculates the confidence level of pseudo supervised loss based on KL divergence,and adjusts the weight of pseudo supervised loss according to the confidence level,enabling the network to learn higher quality pseudo labels,thereby achieving the goal of optimizing network performance.The main idea of this method is to provide an additional reference object for the network,so that the network can determine whether the generated pseudo labels are worth learning.The experimental results show that this method has satisfactory segmentation accuracy.(3)Heterogeneous cross pseudo supervised network based on confidence evaluation.Firstly,construct a heterogeneous cross pseudo supervised network as described above.This article designs a confidence evaluation module in the output section of heterogeneous cross pseudo supervised network networks,which is used to evaluate the quality of pseudo labels generated by two networks and obtain the confidence level of pseudo supervised losses.Taking confidence as a part of regularization and a part of the overall loss of the network,the learning efficiency of the network is improved.The purpose of this method is to solve the problem of unstable quality of pseudo labels produced by CPS when using networks with different systems for cross learning.The core idea is that when the pseudo labels produced by Unet and Swin Unet are similar,the maximum probability of the quality(accuracy)of the pseudo labels is high.The experimental results show that this method performs well and is suitable for heart image segmentation.