Research On Cardiac MRI Segmentation And Quantitative Analysis Based On Fully Convolutional Neural Network

Automatic segmentation and quantization methods for cardiac magnetic resonance image(MRI)can automatically delineate the contour of the anatomical structure of the heart from complex cardiac MRI,accurately measure the functional and morphological parameters of the heart,realize the quantitative analysis of cardiovascular diseases,and improve the accuracy and efficiency of doctors in diagnosing cardiovascular diseases.Therefore,it has become a research hotspot in the field of medical image processing.It is difficult for the traditional cardiac MRI segmentation algorithms to meet the needs of hospitals to process a large number of cardiac MRIs,due to the low automation,computational efficiency,and robustness.By comparison,the fully convolutional neural networks(FCN)based cardiac MRI segmentation algorithms can extract the multiscale and hierarchical semantic features from MRI,enhancing the robustness to the noise,the intensity heterogeneity,and the morphological variability of the anatomical structure.Moreover,the end-to-end structure of the FCN model improves the utilization of data and simplifies the design process of the model.Therefore,FCN has become the mainstream trend in cardiac MRI segmentation algorithms.This dissertation mainly focuses on the left ventricular(LV)segmentation,biventricular(left and right ventricles)segmentation,left ventricular 17-segments segmentation,and quantitative analysis of dynamic cardiac MRI.The main research work is as follows:1.An automatic cardiac MRI processing method based on FCN for left ventricle segmentation is proposed.To improve the segmentation effect on the base and apex of the left ventricle,a novel dynamic pixel-wise weighting strategy is developed to adjust the weight value of each pixel of the training image,guiding the full convolution neural network model to focus on the pixels that are difficult to classify.In addition,a post-processing algorithm based on convex hull algorithm and morphological processing method is proposed to restore the ring structure of some left ventricular segmentation results for accurately extracting the contour of endocardium and epicardium.2.An automatic biventricular segmentation algorithm for cardiac MRI is presented based on a two-stage FCN model.The FCN in the first stage is employed to detect the center point position of the left and right ventricles in cardiac MRI.Then the location information of center points is used to extract the region-of-interest of cardiac MRI and utilized as the prior information to guide the FCN in the second stage to segment the right and left ventricles near the center point.It can effectively solve the under-segmentation and over-segmentation problem at the apex of the left ventricular for the existing algorithms.In addition,a loss function based on the evaluation metrics of false positive and false negative rates is proposed to improve the segmentation performance of the FCN model.3.An automatic left ventricular 17-segments segmentation system for cardiac MRI is constructed based on the standard anatomical model of left ventricular 17-segments and a two-stage FCN model.Firstly,the anatomical keypoints of cardiac MRI were located using the FCN model in the first stage.Then the cardiac MRI was divided into four levels according to the location information of detected keypoints.Subsequently,the MRI slices in different levels were inputted into the specified FCN to segment the left ventricle,reducing the negative influence of anatomical shape and size variability.Finally,the left ventricle segmentation results were segmented into 17-segments based on the polar coordinate transformation and the standard anatomical model of left ventricular 17-segments.4.An automatic quantification system of ventricular functional parameters is developed based on the proposed cardiac MRI segmentation algorithms.Firstly,the left and right ventricles,as well as the left ventricle 17-segments segmentation results,are obtained using the proposed cardiac MRI segmentation algorithms.Then,the ventricular end-diastolic and end-systolic time points are identified by analyzing the curves of left ventricular volume against time.Finally,the ventricular functional parameters,wall thickness,and wall motion are measured from the segmentation results using the calculation formula of ventricular function parameters and Euclidean distance transformation for quantitative analysis of cardiac MRI.The proposed algorithms are tested on multiple public data sets for evaluating the performance.Experimental results show that the above algorithms can effectively solve the existing problems of cardiac MRI segmentation and realize the efficient and accurate quantitative analysis of cardiac MRI.