Research On Intelligent Heart Segmentation Algorithm For Cardiovascular Diseases

In recent years,cardiovascular disease has become the leading cause of death for people around the world,accounting for one third of deaths worldwide.In current clinical care,whole-heart segmentation is an important step for physicians in the diagnosis and treatment of cardiovascular disease.However,whole-heart segmentation still faces many challenges due to the scarcity of training samples,the variability in the shape of heart structures and the ambiguity of the boundaries between different substructures.In this context,this thesis conducts a study of intelligent heart segmentation algorithms for cardiac CT images which are widely used in the medical field,and proposes a cascade framework CM-Tran Ca F with cross-modality domain transfer learning for whole-heart segmentation.The main work is as follows:(1)This thesis conducts a study of the whole-heart segmentation algorithm and proposes the U-shaped Multi-Attention Network MAUNet.In order to solve the problem that a large number of background pixels affect the segmentation accuracy,this thesis add the attention gates AGs at the skip connection of MAUNet to suppress the activation of background pixels.Secondly,to solve the problem of blurred boundary between substructures,this thesis add the position attention block PAB at the bottom layer of MAUNet to aggregate similar features.Finally,in order to utilize the anatomical information between different substructures of the heart,this thesis uses the spatial configuration network SCN to fine-tune the segmentation results.(2)This thesis conducts a study of the modality transfer algorithm for 3D medical images and proposes the Modality-Transfer Network MTN.Then cascading MTN with MAUNet and proposing a cascade framework CM-Tran Ca F with cross-modality domain transfer learning for whole heart segmentation.To make full use of the multi-level information provided by different modality data,this thesis considers the simultaneous segmentation of heart tissue using CT and MRI images.MTN is proposed to unify the data distribution of different modality images,transfer MRI images from MRI domain to CT domain,and apply the data of different modalities to the segmentation of cardiac CT images to expand the training set and further improve the whole heart segmentation accuracy.(3)An intelligent assisted diagnosis system for cardiovascular diseases is implemented.The implementation of this system is based on the CM-Tran Ca F algorithm framework proposed in this thesis,which supports users to browse cardiac images online and can retrieve the trained models through a remote server to obtain the segmentation results of cardiac images instantly and display them through a visualization interface.The value of this thesis’ s algorithm and its application prospects are confirmed from the clinical level.Finally,the performance of CM-Tran Ca F is evaluated on the 40 test datasets provided by the MM-WHS challenge.The average Dice index of this experiment is 0.911,the average Jaccard index is 0.837,and the average Hausdorff distance is 14.386.The segmentation accuracy of the right atrium,left ventricular myocardium and pulmonary artery is the highest so far.