Studies On Left Ventricle Myocardium Segmentation Of Cardiac 4DCT Images Based On Registration

With the improvement of the CT technology,cardiac 4DCT has played important roles in heart diseases diagnosis,images of which can reflect the position and shape of the heart during the entire heartbeat cycle.The abnormalities in the shape and movement of the left ventricle of the heart are important bases for the diagnosis of heart disease while the left ventricular myocardium(LVM)segmentation is a key step in such a diagnosis strategy.Therefore,segmenting the LVM out from the 4DCT is of great merit in studying left ventricular function.Due to the unclear boundary between the LVM and its surrounding tissues and also due to a small proportion of the LVM in the entire image,the segmentation of the left ventricular myocardium from the heart 4DCT image is a challenging task and most traditional segmentation methods cannot achieve satisfied results.Therefore,this thesis devotes itself in developing segmentation algorithms based on atlas and image registration specifically according to the characteristics of left ventricular myocardium in 4DCT images of the heart.The main work and contributions are as follows:According to the fact that the LVM occupies only a small part of the entire 3DCT image and global registration algorithms cannot highlight such a region of interest(ROI)in the registration process,a two-stage registration based segmentation method is proposed respectively for ROI positioning and finer LVM segmentation.In the first stage,a parametric total variation non-rigid image registration method is used to perform global and coarse registration to align the segmented 0%phase atlas in 4DCT to other phase images whose ROI region containing the LVM can then be determined upon mapping segmentation labels from the segmented atlas to the one to be segmented.In the second stage,a finer registration is performed with the same registration method in between the above extracted ROI pairs in the 0%atlas and in other phase images to map segmentation labels from atlas ROI to the ROI in other phase images to obtain the final segmentation result.Experimental results prove that the algorithm proposed in this thesis can effectively improve the segmentation accuracy of left ventricular myocardium after adding a small amount of calculation timeIn addition to the above method,a left ventricular myocardial segmentation method is also proposed based on the convolutional neural network(CNN)registration and corrective learning.In this method,the spatial resolution of the image is first down sampled to reduce the computational cost and storage requirements of the neural network Then,a weakly-supervised CNN is trained in the low-resolution space of the image to perform registration between the segmented 0%phase image in 4DCT and other phase images.Upon introducing edge information extracted with six 3D Sobel operators to the loss function of the convolutional neural network,edges in the segmented results are forced to approach those in the golden standards step by step during the iterative training process of the network,which drives the network to achieve higher segmentation accuracy.Subsequently,the segmentation result is scaled back to the original resolution and the final segmentation result is obtained after correcting all system errors caused by resolution reduction and other factors with a corrective learning method.Experimental results show that,upon introducing the edge information in the network,our proposed method outperforms other network models without edge information in segmentation accuracy.Even if the registration is performed in lower spatial resolution,we can still achieve a segmentation accuracy similar to that obtained in higher resolution upon performing a corrective learning procedure.And reducing the resolution of the image can also reduce the memory requirements and improve the efficiency of the algorithm.