Semi-supervised Cardiac Segmentation Of MRI Based On Transfer And Contrastive Learning

In recent years,cardiac diseases have gradually become one of the major threats to human health.According to the statistics of the World Health Organization,one out of every three cases of death due to cardiac diseases.Therefore,the accurate evaluation of cardiac function is of great significance for the prediction and diagnosis of cardiac diseases.As a non-invasive cardiac examination technique,cardiac MRI(magnetic resonance imaging)can directly reflect the structural characteristics of the heart,and has been widely used in medical clinical practice.Accurate segmentation of cardiac MRI images can provide doctors with anatomical structure information needed for disease diagnosis and treatment,and help doctors accurately evaluate cardiac function.Most of existing deep learning methods adopted the fully supervised training mode,and the model performance depended on a large number of labeled samples.However,the cardiac structure is complex,and the labeled MRI samples usually need to be manually drawn by experienced doctors and experts,and the labeling process is easily disturbed by various factors,resulting in a small scale of high-quality labeled samples,which seriously restricts the performance of fully supervised learning models.To solve the problem of lack of labeled samples at present,this thesis proposes a semi-supervised cardiac MRI segmentation framework based on transfer and contrastive learning.The main contents are summarized as follows:(1)A contrastive learning framework(CPCL-Net)with dual-coded and single-decoded structure is proposed for cardiac MRI segmentation,including a major encoder(ME),an auxiliary encoder(AE)and a share-decoder(SD).CPCL-Net uses MT-Unet as the backbone network,which has recently been applied in the field of medical image segmentation,adds attention mechanism to the U-Net network and improves the expression of feature structure information.In order to be more suitable for the cardiac segmentation task,this thesis also designs the joint pixel contrast loss,so that CPCL-Net can directly conduct feature learning at the pixel level.Through constraint training from the pixel level and image level,the model is optimized,the potential diversity of feature representation and the segmentation accuracy are all further improved.In the experimental part,the advantages of CPCL-Net and the effectiveness of joint pixel contrastive loss are verified.(2)Although the CPCL-Net proposed in this study has achieved good segmentation accuracy,the accuracy has not improved much.Therefore,in order to further improve the segmentation accuracy of CPCL-Net,this thesis designs a dynamic adaptive update module,which can dynamically generate reweighted update factors ,.The adaptive reweighting factor can adjust the training contribution of sample ,while the factor is modeled by a spatial attention module to emphasize the contribution of pixels within the feature map.The module allows the CPCL-Net to focus on the training contribution of samples and the contribution of spatial pixels simultaneously.By using these two factors to assist in the adaptive momentum updates of the network and adjust the loss function of the main encoding network,the training weights of difficult negative samples can be adjusted from easy to difficult.This focuses the training on rare and difficult negative samples,enhances locally relevant features,suppresses irrelevant features,and ultimately enhances the ability to distinguish high-level semantic information.Experimental results demonstrate that the dynamic adaptive weighting module can effectively strengthen edge segmentation of details,significantly improving the cardiac segmentation performance of CPCL-Net.