Cardiac MRI Image Segmentation Method And System Based On Level Set And Convolutional Neural Network

Heart disease is one of the major causes of unnatural death in humans at present.The research and treatment of cardiac pathology often rely on a large number of cardiac imaging methods.As a common used technology,MRI provides important information for diagnosis and treatment of heart disease.Many clinical parameters are obtained through segmentation of the ventricle from MRI images..Clinical segmentation is usually performed manually by experts,which is not only time-consuming and labor-intensive but also has high inter-observer variability.Therefore,developing fast,accurate,repeatable and fully automatic algorithms of ventricle segmentation are of great significance in study of the heart and its diseases.Although researchers have proposed a large number of segmentation algorithms,the segmentation of ventricular regions is still very challenging.In this paper,we study algorithms of left ventricular segmentation systematically.The specific work and main contribution are as follows:(1)A left ventricular segmentation method is proposed based on the region locating algorithm and the convexity-preserving level set.Firstly,each MRI image is divided into two clusters,background and foreground,using the SBGF-new SPF method.Secondly,each connected region is assigned with a different connected component label and an initial contour of the ventricle is obtained based on the improved region selection algorithm.Then,a contourconstrained level-preserving level set and a rounding-level set algorithm are combined to control evolution of the endocardial contour.Finally,the epicardium is extracted using the convexity-preserving level set method based on the constraint of mean intensity with the contour initialed by the segmented endocardium.Experimental results show that the proposed algorithm are more automated and accurate with fewer iterations compared with other advanced segmentation methods.(2)A segmentation model is proposed by combining the convolutional neural network and the level set for extraction of the left ventricular from MRI images.The model contains multiple expansion rate dense layers,level set layers,and fusion layers.On one hand,the dense convolutional neural network is constructed to learn image features and output a probabilistic segmentation result which is used to initialize the contour of the level set layer.On the other hand,the loss function of the dense neural network is updated using the output of the level set layer and image features,which results in updating parameters of the dense network.The above two methods interact and influence each other for the collaborative learning in order to improve segmentation accuracy is improved.Experiments prove that this proposed method is more robust and accurate than the single convolutional neural networks or the level set method.(3)A left ventricular image segmentation and analysis system is designed and established.The system contains three major modules: image data processing module,image segmentation algorithm module and segmentation result evaluation module.Besides,the system provides a variety of functions such as data processing of left ventricular images,segmentation of endocardium and epicardium,and evaluation of segmentation results.