Research On Segmentation And Feature Extraction From Three Dimentional Carotid Ultrasound Images

Cerebrovascular disease has become the leading cause of mortality and morbidity in the world,and carotid atherosclerosis with plaque rupture is the main cause of cerebrovascular disease.Major cerebrovascular events may occur in patients without known preexisting symptoms,thus it is important to monitor the progression and regression of carotid plaque for evaluating the therapy patient's response to therapy.The aim of this paper is to develop automatic methods to measure palque burden and extract plaque characteristics from three-dimensional ultrasound(3D US)images to monitor plaque progression and regression for evaluation of drug treatment effect.Medical imaging is the main tool in clinic for carotid atherosclerosis examination,and ultrasound imaging is the most common method because of its low-cost,convenience and non-invasive nature.However,the quality of 2D ultrasound images varies from the experience of operators.3D ultrasound overcomes these weaknesses and provides a more reliable,convenient,repeatable and effective carotid atherosclerosis diagnostic technique.Some studies showed that 3D measurements of plaque burden,such as carotid artery wall volume(VWV)and carotid plaque volume(TPV),have shown to be more sensitive to the changes of plaque burden than one-/two-dimensional measurements,and have strong related to the risk of cerebrovascular events.Also,the changes of texture features extracted from three-dimensional carotid ultrasound can reflect the changes of carotid plaque components,and the roughness of plaque surface can descibe the plaque surface ulcer,both of which are important factors for judging plaque vulnerability.Therefore,in this paper,we will study the automatic carotid VWV and TPV segmentation methods and the carotid plaque feature extraction methods from 3D US images.Firstly,the dissertation proposed a VWV segmentation method from 2D transverse slices of carotid 3D US images.Since training the three-dimensional convolution neural network(CNN)requires a lot of 3D US images,it is difficult to obtain so many carotid 3D US images to train the 3D CNN in the early stage of the research.Therefore,we sliced the3 D US images to 2D images along the transverse view of carotid artery and used these 2D images for Lumen-intima boundary(LIB)and Media-adaventitia boundary(MAB)segmentation.A dynamic convolutional neural network(Dynamic CNN)was proposed to segment the MAB contours.The improved U-Net was applied for LIB segmentation,which allows the network to be trained end-to-end for pixel-wise classification.The results showed that our method reuced the processing time and improved the segmentation accuracy for both MAB and LIB.Then,a voxel-based fully convolution nerual network(Voxel-FCN)was proposed to segment VWV from carotid 3D US images automatically.Unlike the general FCN,we used 3D pyramid pooling module(PPM)to extract the context relationship between objects in the encoding path of Voxel-FCN,and an attention mechanism was added to the connection modulein the decoding path,which allowed the network to focus on more important features.The results showed that the Voxel-FCN method greatly reduced the processing time with a promising accuracy.Thirdly,a plaque automatic segmentation algorithm was proposed based on a small number of labeled samples to obtain carotid TPV measurements.In this paper,an unsupervised encoder-decoder network was proposed to be trained with unlabeled images,and then the model was used to initialize parameters of the improved U-Net network and fine-tuned all the parameters in the model using labeled images.Besides,a Monte Carlo Dropout is proposed to introduce uncertainty into the network to improve segmentation accuracy and reduce over-fitting.The experimental results show that the proposed method had higher segmentation accuracy than the existing plaque methods and the original U-Net.Finally,we researched on the feature extraction methods of carotid plaque from 3D US images.The VWV and TPV measurements were obtained using methods introduced in the previous three chapters.A 3D co-occurrence matrix features was proposed to describe the texture features of plaques.Moreover,a feature description method based on three-dimensional fractal dimension was proposed to describe the irregularity inside plaques and the surface roughness of plaques.The above features were combined to obtain a stronger plaque feature descriptor and the clinical data was used to evaluate the discrimination ability between patients treated with atorvastatin and placebo.The experimental results showed that the 3D fractal dimension features could be used to describe the changes of plaque.Moreover,the optimized combination of features showed a significant difference in distinguishing patients treated with atorvastatin and placebo.Thus,in this dissertation,the author proposed three carotid 3DUS image segmentation methods to generate the VWV and TPV measurement and studied the plaque feature extraction methods to monitor plaque progression and regression in use of the evaluation of drug treatment effect.