Research And Realization Of Carotid Artery Ultrasound Image Plaque Detection

Cardiovascular disease（CVD）caused by carotid artery（CA）disease is one of the most common causes of death after heart disease and cancer.This disease prevents blood from flowing from the heart to the brain and face,leading to lack of oxygen to the brain and triggering sudden illness.Atherosclerosis is the pathological basis of cardiovascular disease,and early detection and diagnosis of carotid plaque is critical to the later treatment of patients.Ultrasound imaging is capable of rapid imaging and has the advantages of being non-invasive,radiation-free,simple,reproducible and affordable.Clinical studies have shown a high agreement between carotid plaque detection using ultrasound images and detection by pathology,so ultrasound is usually chosen to examine carotid arteries.However,the resolution of ultrasound images is generally low and the detection method is usually manual or semi-automatic,making the results subjective and errorprone.This paper conducts a study based on carotid ultrasound images to solve the problems of ultrasound sequence key frame extraction,ultrasound image super-resolution,carotid lesion region segmentation and carotid plaque detection,which provide an effective method for clinical computer-aided diagnosis of carotid artery diseases.The work and innovation points of this paper are as follows.（1）During the ultrasound probe scanning,noise,uneven grayscale pixels and doctor’s scanning technique lead to the problem of more redundant information in ultrasound sequences.this paper proposes a carotid ultrasound sequence key frame group extraction algorithm to address this problem.Firstly,the information frames containing the cross-section of blood vessels in the carotid ultrasound video are selected by combining SVM and HOG features of blood vessels;Secondly,the ultrasound video information frames are clustered according to the shape of blood vessels;Finally,the image quality evaluation algorithm NIQE-K is used to select the best quality frames from the information frame cluster as the final key frame set for subsequent processing.Quantitative indexes PLCC,SROCC,RMSE,and KROCC are used to evaluate the experimental results and verify the effectiveness of the algorithm.（2）Ultrasound images are easily restricted by ultrasound instrument models,imaging principles,and manual operations,resulting in low image quality and poor resolution,while the lesion area is small,making disease detection difficult.A super-resolution reconstruction algorithm for ultrasound images based on generative adversarial networks（USRGAN）and region of interest extraction algorithm for ultrasound images are proposed.USRGAN adopts multi-scale convolutional module and a network structure of multi-level image feature fusion.The model is trained with rabbit liver,human liver,carotid artery and other ultrasound image datasets.The subjective and objective quality of the reconstructed ultrasound images are evaluated,and the experimental results show that USRGAN performs better than Bicubic,SRCNN and SRGAN methods in each dataset,and is able to clearly reconstruct lesion details and organ structures.In the region of interest extraction,an algorithm for automatic localization of region of interest extraction combining vascular morphological information and Gaussian mixture model（GMM）is proposed,which taking advantage of the bright and clear characteristics of distal vascular imaging and obvious inner and outer membrane delamination.The ablation experiments showed that the region of interest extraction improved the plaque classification accuracy.（3）To address the problems of low accuracy and unclear classification of traditional carotid plaque detection algorithms,a plaque detection model（DCP）based on convolutional neural network is proposed in this paper.The distal carotid vessel wall region of interest is used as the training set to achieve the classification of carotid plaque.The experimental results show that the accuracy,sensitivity and specificity of DCP model are 91.5%,89.8% and 93.9% respectively,which are better than the traditional algorithm and have better robustness.Meanwhile,a plaque property detection model（PPD）based on convolutional neural network is designed by combining the attention mechanism,which solves the dichotomous problem of determining the vulnerability and stability of carotid plaque properties.The accuracy of the model reaches 85.5%,which is 4.2%higher than the accuracy of the backbone network.The proposed DCP model and PPD model in this paper provide an effective exploration for clinical computer-aided diagnosis of carotid artery disease.