Research On Ventricular Image Segmentation Method Based On Nn-UNet

In recent years,cardiovascular disease has become the leading cause of annual mortality in China,and the demand for prevention and diagnosis of cardiovascular disease has also been increasing year by year.Accurate segmentation of the left and right ventricles in cardiac magnetic resonance imaging can help quantify and analyze cardiac function.In the past,this segmentation process required manual operation by professional imaging physicians,which was time-consuming and prone to manual errors,Therefore,it is particularly important to develop a fast,accurate,automatic,and efficient heart MRI image segmentation method.With the continuous progress of artificial intelligence technology,especially deep learning technology,deep learning related heart MRI image segmentation methods have begun to demonstrate excellent performance.Based on nn-Unet,a network framework,which can automatically generate hyperparameter for most training based on a specific dataset,this paper improves the network baseline used by the framework,optimizes the post-processing steps and loss function and other factors that have some impact on the experimental results,and thus proposes an improved version of nn-Unet,The results on the dataset show that this scheme has excellent performance in left and right ventricle and myocardial segmentation tasks in cardiac magnetic resonance imaging.Regarding the improvement plan,this article first improves the most basic nn-UNet to make it more suitable for the segmentation task of cardiac magnetic resonance images.The optimization methods mainly include: improving the downsampling process,using Res Net module to replace the general convolution module in the encoder.This method can effectively reduce the number of network parameters while increasing network depth,thereby improving network performance;To improve the upsampling process,Axial attention mechanism modules were introduced in some decoders to effectively filter the feature information that requires focused learning in the transmitted feature maps.In addition,aiming at the problem of class imbalance in the dataset,we adjusted the loss function used in the network framework,and used a structure generated by combining a small batch weighted Dice loss function and a spatially weighted cross entropy loss function as the loss function used in the experiment,which effectively alleviated the problem of class imbalance in the image and improved the performance of the model in the segmentation of left and right ventricle in cardiac magnetic resonance imaging.Regarding the data used in the experiment,a internally collected dataset was jointly established by the laboratory and the collaborating hospital.Multiple experiments were conducted on this dataset and the ACDC(Automated Cardiac Diagnosis Challenge)dataset to verify the excellent segmentation performance of the proposed network framework and the effectiveness of the improvements made to the basic nn-UNet framework in this paper.The ACDC dataset includes 100 annotated data cases,and the self collected dataset of cooperative hospitals includes 80 annotated data cases.The average Dice coefficients of the corresponding segmentation results for left ventricle,right ventricle,and myocardium on the ACDC dataset are 0.967,0.924,and 0.911,respectively.The average Dice coefficients of the corresponding segmentation results for left ventricle,right ventricle,and myocardium on the self collected dataset of cooperative hospitals are 0.949,0.926,and 0.903,respectively.In order to further improve the effect of segmentation,a Connected Components algorithm was introduced in the post-processing stage to suppress false positives and missegmentation phenomena that appeared in the heart MRI image segmentation experiment.The results showed that this method effectively reduced the false positives and missegmentation phenomena in the model inference process.The experimental results show that the addition of this algorithm has a very good effect on reducing the Hausdorff distance,reducing the average Hausdorff distance by 4.323 mm and 3.941 mm on the ACDC dataset and the self collected dataset of cooperative hospitals,respectively.