Deep Learning Based Research On Cardiac Medical Image Segmentation

The rates of mortality and morbidity associated with cardiovascular diseases have increased over the recent years all of the world;these diseases are regarded as the primary cause of death in many countries including China.The early quantitative diagnosis and risk assessment of heart diseases is crucial for preventing sudden death and the improving the quality of life of patients.Computed tomography(CT),magnetic resonance imaging(MRI)and other imaging technologies are clinically applied to perform cardiac imaging.Cardiac images are segmented with manual methods that yield precise results,but involve poorly reproducible and time-consuming steps.Moreover,manual segmentation cannot process a large amount of imaging data to diagnose a heart disease quantitatively.These limitations of manual segmentation above have prompted researchers to develop semi-automatic or fully-automatic segmentation methods for cardiac image analysis.However,the development of accurate cardiac segmentation techniques remains challenging,because of many limitations such as the complexity of the cardiac structures,the fuzzy substructure boundaries and motion artifacts and noise due to heart beat during image acquisition.Whole heart segmentation refers to the extraction of the volume and shape of the heart substructure.The heart substructure specifically includes four chambers(left ventricle,right ventricle,left atrium,right atrium),left ventricle myocardium,and large vessels(ascending aorta,pulmonary artery).In this dissertation,CT cardiac medical images are used for the whole heart segmentation research in the field of deep learning.We propose two heart segmentation methods based on 3D ML-UN network and 3D DML-UN cascaded network respectively,and have achieved automatic whole heart segmentation of 3D medical images.The main content of this thesis include:(1)This thesis proposes a 3D U-net-based heart segmentation and ROI detection algorithm,which is used as a preprocessing part of the tast of whole heart substructure segmentation and realizes the ROI detection of the original image.The segmentation results of this algorithm are considered as the basis for sub-structural segmentation of the whole heart.(2)This thesis proposes a 3D ML-UN network based on ML guidance unit for whole heart segmentation,which improves the performance of the classical U-net lacking use for multi-scale.So,the network training can be guided by multi-scale outputs and losses making the network more instructive and the segmentation more accurate.(3)This thesis proposes a cascaded 3D DML-UN network framework based on 3D ML-UN in whole heart segmentation.The cascaded network framework consists of a coarse network and a fine network.The coarse network realizes a rough segmentation of the whole heart,and the fine network realizes the refinement of the heart segmentation based on the rough segmentation results.The experimental results show that the method proposed in this thesis greatly improves the accuracy of segmentation of the left ventricle and ascending aorta.On the experimental data set,the proposed method obtains the best average segmentation accuracy compared to the currently published methods and it can be used for whole heart segmentation with high accuracy.