Research On Segmentation And Quantitative Analysis Methods Of Heart Images

Cardiovascular Diseases(CVDs)has become one of the major diseases that seriously threaten human health and life safety in Asia.With the continuous development of medical technology,different types of medical imaging equipment are gradually used in the diagnosis of cardiovascular diseases,such as magnetic resonance imaging(MRI),computed tomography(CT),positron emission tomography Imaging(PET),etc.Left ventricular(LV)segmentation and quantitative analysis are one of the important steps in the diagnosis of heart disease based on cardiac images acquired by medical imaging equipment.LV segmentation refers to the precise division of the left ventricular boundary on different cardiac images,including the boundary of the endocardium and the epicardial boundary.Quantitative analysis of the left ventricle refers to the estimation of the LV wall thickness includes eleven indicators,the area of the left ventricle(the area of the heart chamber and the area of the left ventricle),the regional wall thicknesses(anterior,anterospetal,inferoseptal,inferior,inferolateral,anterolateral)and the diameter of the left ventricle(the diameter of the three fixed directions).LV segmentation helps radiologists obtain the boundary of LV myocardium and provide visual aid to the radiologist.Currently,clinicians LV segmentation mainly by manual segmentation of radiologists.However,it is a time-consuming,tedious and inefficient job.The data shows that it average takes 20 minutes for a clinician to segment cardiac MR images of one patient(20 subjects).In clinical practice,a large number of patients and different imaging devices will produce a large number of cardiac images.It is a time-consuming work for radiologists to segment these images.Even more improtant,the individual differences between patients,the shape changes of the myocardium caused by cardiac exercise,the presence of papillary muscles,the noise of different modal images,and so on will affect the accuracy of manual segmentation of radiologists.Besides,due to the inevitable interference of personal subjective factors,the result of cardiac segmentation is inevitably lacking certain objectivity.LV quantification provide a better understanding of the position,boundary and thickness ofthe myocardium in patients,which is of great significance for the diagnosis of cardiovascular disease.The existing left ventricular quantitative analysis method mostly relies on the radiologist to segment the left ventricle.Based on the results of manual segmentation,the estimated left ventricular area,ejection fraction and wall thickness are measured.The error of the manual segmentation and the following estimation will affect the accuracy of the LV quantification.Meanwhile,it also has a higher time cost,no matter the manual segmentation,but also the following estimation.This leads to the direct left ventricular manual segmentation error and affect the results of quantitative analysis of the left ventricle.At the same time,based on the large amount of time consumed by left ventricular segmentation,further measurement estimates also require a significant amount of time.Therefore,manual left ventricular quantitative analysis not only has secondary errors,but also has a higher time cost,which is not conducive to the formation of modern computer-aided diagnosis.In order to improve the efficiency of cardiac image segmentation and quantitative analysis,free radiologists from lengthy and inefficient work,computer-based cardiac segmentation and quantitative analysis methods are imperative.Based on the above analysis,this paper proposes a series of left ventricular segmentation and quantitative analysis methods:(1)This paper proposes a Model Adaptation-Shape regression method for LV segmentation.Current LV segmentation algorithms are based on single-modality or multi-modality images.The generalization ability based on the single-modal LV segmentation algorithm is relatively po or,and it is often necessary to retrain when applied to the new modality.Multi-modality LV segmentation algorithms often require images of different modalities of the same patient.Considering the similarity of the shape of CMR images and CT images,we first proposed the mixed modality images,which is different from the common multi-modality images.It refers to the image containing multiple modalities in the image,but the images of multiple modalities are from different patients.The Model adaptation-Shape regression method makes full use of the similarity of the left ventricle shape in different cardiac images.It has high generalizationperformance and do not require secondary training for the new modal image segmentation.(2)This paper proposes a semi-supervised GAN for left ventricular segmentation.A large number of training images can often improve the accuracy of the algorithm.However,image annotation is an extremely cumbersome task with a lot of work.Especially for medical image annotation,inexperienced radiologists are often difficult to mark correctly.Therefore,in order to make full use of a large amount of unlabeled data,this paper proposes a semi-supervised GAN for left ventricular segmentation.(3)This paper combines deep neural network with ensemble learning,and proposes a deep ensemble learning network to realize quantitative analysis of left ventricle.Quantitative analysis of the left ventricle refers to the estimation of various indicators of the left ventricle.In the diagnosis of cardiovascular disease,accurate left ventricular quantitative analysis results can provide an intuitive and accurate basis for radiologists.Traditional ensemble learning method can effectively improve the accuracy of individual learners.Based on this,this paper designs three sub-networks from three pathes,and obtains the final quantitative analysis results by integrating the results of three networks.This method can make full use of the advantages of left ventricular quantitative analysis on different pathes,and improve the accuracy of estimation.(4)Considering the importance of simultaneous cardiac image segmentation and quantitative analysis,this paper proposes a new deep neural network to achieve cardiac image segmentation and quantitative analysis.In the existing cardiac image analysis work,the heart image segmentation work often only gives the segmentation result,and the quantitative analysis result by the segmentation result requires secondary calculation,and it is difficult to avoid the error in the secondary calculation process.The direct left ventricular quantitative analysis work often fails to give the segmentation results,and it lacks certain visual aids for the clinical diagnosis.Therefore,this paper proposes a new deep neural network to achieve cardiac image segmentation and quantitative analysis,and to provide accurate analysis results,while providing visual aid to radiologists,so as to better achieve computer-aided diagnosis of heart disease.To test and evaluate the performance of the left ventricular segmentation and quantitativeanalysis methods presented in this paper,we conducted a series of experimental comparisons.The experimental results show that the proposed method has high accuracy in cardiac segmentation and quantitative analysis.It reveals its effectiveness and potential as a clinical tool.