Automatic Cardiac Bi-Ventricle MRI Segmentation And Disease Classification Using Deep Learning

Cardiac MRI offers key information for diagnosis by evaluation of the function and structure of the cardiovascular system.Segmentation is an important step for calculating clinical cardiac indices such as myocardium thickness,volume of left ventricle and right ventricle,ejection fraction and so on.Manual delineation by experts is currently the standard clinical practice.However,manual segmentation is tedious,time consuming,relying on doctor's subjectivity and experience.Human hearts have different shapes and sizes and these pathological features are various.The segmentation accuracy of conventional methods such as pixel classification and random tree are limited and far from clinical application.Meanwhile,existing methods based on deep learning mostly singly segment left ventricle or the right ventricle.Although there are also some researches on dual LV and RV segmentation,but the accuracy still needs improvement.Therefore,cardiac MRI segmentation is still a challenging task and important in the field of medical image analysis.MICCAI 2017 ACDC database was used in our research and the data is divided into 5 groups according to pathology: dilated cardiomyopathy,hypertrophic cardiomyopathy,previous myocardial infarction,abnormal right ventricular and normal subjects.In this work,we performed object detection using a YOLO-based network and get region of interest from a whole sequence of diastolic and systolic MRI.Then,automatic cardiac bi-ventricle MRI segmentation was constructed using fully convolutional neural networks(FCN),and a pixelwise segmentation mask automatically was obtained by feeding ROI in to our net.The left ventricular,myocardium,and right ventricle were segmented at the same time.The FCNbased network combined the fine low layer information with coarse and high layer information after deconvolution.Finally,14 features were extracted,including 12 features based on image information using the segmentation results and the two characteristics of the patients' information.the heart disease classification of patients was done using two methods: XGBoost and full-connected network.And we made comparison between these two methods according to the classification results.The speed of YOLO-based heart detection was 49 frames per sec,average precision was 72.6.The FCN-based cardiac MRI segmentation worked well,and even basal slices,left cavity,myocardium and right ventricle can also be well segmented.The segmentation results showed that the proposed method reaches scores in sensitivity: 0.891,specificity: 0.976,Dice coefficient: 0.889,Jaccard coefficient: 0.964,Hausdorff distance: 10.275 mm.The segmentation performance of left ventricular myocardium was better than the right ventricle.the accuracy of the heart classification using XGBoost was 88%,which shows relatively robust performance,and the accuracy of the heart disease classification based on the full connected network was 90%.This method is easy to understand and implemented in theory but it is not as robust as XGBoost.Compared with the single segmentation of left or right ventricular,biventricular segmentation is more in line with the clinical needs.Cardiac MRI biventricle segmentation will be a research hotspot in the field of medical image analysis.