Research On Image Segmentation Algorithm Based On Deep Neural Network And Its Application In Ventricular And Coronary Calcification

With the advancement of medical career,a large number of medical image data have appeared in clinical practice.How to rapidly process these medical images to promote the diagnostic efficiency has always been a problem faced by researchers and medical clinicians.At present,cardiovascular disease is the first cause of death in the global population.Therefore,research on clinical cardiac images has always been a hot topic.In recent years,researches on cardiac images include segmentation of cardiac chambers,coronary artery calcification,etc.for medical images of various imaging modalities.Subsequently,on the basis of these segmentation results,diversity cardiac function indices are calculated.In clinical practice,the objective organs are routinely segmented by clinicians manually.However,manual segmentation is time-consuming,tedious,inefficient and inconsistent.Therefore,the studies on automated segmentation method have gradually wake the attention of researchers.At present,the automated segmentation methods are mainly divided into traditional techniques-based methods and deep learning-based methods.The methods based on traditional techniques,which include threshold-based methods,graph cuts-based methods,etc.,generally seeks to detect the boundary between foreground and background.However,for clinical patients,it is common for objective organs to occur pathological changes and diverse appearance.In result,the complex anatomical structures make these images more challenging for these traditional methods.For deep learning-based segmentation methods,they commonly train a deep neuron network model using a large amount of images.The model classifies pixels of foreground and background by rich object information to realize organ segmentation.The study aims to the segmentation of cardiac bi-ventricle and coronary artery calcification based on deep neural networks.It includes the following aspects:1.The paper introduces the importance of segmentation for diagnosis of cardiovascular disease via illustrating the contribution of segmentation for clinical cardiac indices,such as mass and volume of bi-ventricle and coronary artery calcium scores.Subsequently,it lists the relevant basis and theories of image segmentation methods.Then it briefly combs the clinical methods and the commonly used classical segmentation methods proposed in recent years.And,advantages and disadvantages of these previous methods are analyzed.In view of the shortcomings of traditional methods,this study proposes the automatic segmentation methods based on deep neuron network,and validates them on cardiac bi-ventricle and coronary calcification.2.In the cardiac images,the border between ventricle and background is blurred and indistinguishable,especially in the basal and apical slices.And the shape of endocardium and epicardial contours varies greatly from slice to slice and at different stages.Also,intensity unevenness exists in the images.In view of the challenges of clinical ventricular segmentation methods and the shortcomings of traditional methods,this study proposes a bi-ventricle segmentation method for cardiac magnetic resonance images based on deep belief network.This is the first time that deep belief network is combined with regression to segment bi-ventricle.The method formulates such segmentation problem as a regression task.It uses the local DAISY feature of cardiac images as input,and then learns a regression model of bi-ventricle contour coordinates via deep belief network.Regression model combined deep learning technology and local DAISY descriptor can capture high-level image information and realize accurate segmentation of bi-ventricle with lower computational cost and fewer assumptions.Performance of this method is validated on 145 clinical subjects(2900 cardiac magnetic resonance images in total)from two health care centers.The method achieves high consistency between estimated and manual boundaries on both Pearson coefficient and Dice metric.Therefore,the proposed bi-ventricle segmentation method can be used as an assistant tool for the diagnosis of cardiovascular disease.3.The coronary artery calcification plaques are irregular in shape and varies greatly in volume.Because of intensity and composition of aortic calcification,spine and ribs are close to coronary calcification,the segmentation procedure is greatly affected by the interference from these tissues.In view of the challenges of clinical coronary calcium scoring methods and the disadvantage of previous automatic calcium scoring methods,this study proposes an end-to-end framework for coronary calcification detection based on DenseNet and U-Net.The method uses the DenseNet-based module and U-Net-based module to extract the intra-slice and inter-slice features.Rich calcification features are learned by the joint learning of two modules to achieve accurate artery-specific calcification detection.Non-contrast cardiac computed tomography scans of 169 clinical patients from different centers are used to validate performance of the proposed framework.The results show that the estimated calcification results are in good agreement with the clinical gold standard.