Research On Cardiac Image Segmentation Approach Based On Deep Learning

At present,Cardiac disease not only affects people's quality of life but also poses a serious threat to human health.Therefore,early diagnosis and treatment of heart disease is essential.Cardiac image segmentation can provide effective and reliable decision-making information for doctors in the diagnosis and treatment of heart disease and its pathological analysis,and plays an important role in the processing and analysis of medical images.Since Magnetic Resonance Imaging(MRI)technology can provide high-resolution cardiac soft tissue images in a non-invasive manner to obtain distraction information of the subject's heart.Cardiac MRI images have been widely used to segment cardiac organs,tissues and lesions.Because manual segmentation is subjective,time consuming,inefficient,and the process of segmentation is not repeatable.Therefore,the automation of accurate segmentation of cardiac MRI images is of great clinical significance.However,because of the characteristics of cardiac MR imaging technology(noise,image intensity non-uniformity,etc.),the structural characteristics of the heart target itself(target shape change,target size,etc.)and the difference in heart between different patients,the automated and accurate segmentation from cardiac MRI images still face big challenges,such as the Right Ventricle(RV),the Left Atrium(LA),and the Left Atrial Scar segmentation.According to the characteristics of cardiac MRI images and the needs of clinical application,two heart image segmentation approaches based on deep learning are proposed:1.For the local weak/no boundary of the right ventricle in the cardiac MRI image,the non-uniformity of the pixel intensity inside the ventricle,and the variation of the shape of the right ventricle.This paper proposes an approach of cardiac right ventricular regression segmentation based on convolutional neural network.This approach treats right ventricular segmentation as a regression problem at several points on the right ventricular border.It allows the segmentation of the right ventricle relies solely on the global information of the cardiac MRI image.The approach uses the convolutional layer of the convolutional network to automatically learn the deep global features of the cardiac MRI image,and then use the folly-connected layer to regress the boundary point of the right ventricle based on the learned features.This segmentation method provides a flexible representation of the shape of the heart with points,without regard to the non-uniform and local weak boundary of the internal pixels of the right ventricle.The experimental results show that the proposed approach achieves a high degree of consistency with the clinical expert in manually segmenting the right ventricle of the heart.2)In order to solve the problem of manual intervention for multi-step segmentation of atrial scar and small target scar segmentation,this paper proposes an end-to-end multi-object segmentation approach.It automatically segments the left atrium and its attached pulmonary veins and atrial scars.The approach mimics the operation of the clinician to observe the image,and slices the 3D image in the axial plane to finally achieve 2D segmentation.The approach consists of two parts:a multi-view learning network,which mainly learns the correlation between axial slices through long-term short-term memory models.Meanwhile it incorporates the information of the coronal plane and the sagittal plane to compensate for the 3D image in the process of slicing,which leads to the loss of the spatial information.Based on the enhanced axial slice features of multi-view learning,the segmentation of the left atrium and its attached pulmonary veins is achieved;The attention mechanism model,which can directly study the attention weight distribution map on the original image,thereby enhancing the features of the scar region of the axial slices,thereby enabling the model to focus on small scars during the segmentation process.By comparing with the latest methods,this method achieves the best segmentation accuracy.Based on the segmentation approaches proposed above,this paper achieves the accurate end-to-end segmentation of cardiac MRI images.It can thus provide an effective and accurate segmentation tool for the clinic.This can bring convenience and efficiency to the clinical diagnosis and treatment of heart disease.