Convolutional Neural Network Based Carotid Plaque Recognition Over Small Sample Size Ultrasound Images

Cardiovascular and cerebrovascular diseases have a worldwide distribution,whose morbidity and mortality have been continued to increase in recent decades.Atherosclerosis is the main pathological basis of cardiovascular and cerebrovascular diseases.As the large artery connecting the heart and brain,carotid artery can reflect the development trend of major atherosclerosis in human body.The common method to examine carotid atherosclerotic plaques in clinic is ultrasound imaging technology.The automatic recognition of carotid plaque ultrasound images with computer,no doubt,can greatly improve the efficiency and accuracy of clinical diagnosis.Convolutional Neural Networks(CNNs)have distinguishing advantages in the field of image processing,but the prerequisite for training a CNN is massive sample data set.The acquisition of the gold standard for medical imaging is time consuming and expensive.It is of great scientific value and clinical significance to study how to improve the ability of CNN to identify the ultrasound images of carotid plaques over small samples.In this thesis,three-dimensional(3-D)carotid ultrasound images of 36 subjects,supplied from the Robarts Research Institute in Canada,was used to resample and get 1,828 transverse two-dimensional(2-D)images of the common carotid artery.And three different types of Regions of Interest(ROIs)were extracted based on the intima-lumen interface and media-adventitia interface marked by the physician.The original ROI1 was selected directly.The media-adventitia interface of ROI1 was segmented to get ROI2.And the intima-lumen interface of ROI2 was segmented to obtain ROI3.The experimental sample set was established through the traditional data augment method and a total of 16,452 samples were obtained for each type of ROIs.In the experiment,we first trained a CNN to recognize ultrasound images of carotid plaque.And then,three strategies,including adding new training set samples generated by the Generative Adversarial Network(GAN),transfer learning and active learning,were used to try to improve the recognition performance of the CNN.The results show that the new samples generated by the GAN can fit the contour of the carotid artery well,but lose the important details of intima-media,and have no significant improvement on the recognition performance of the CNN.Active learning requires multiple trainings on the CNN.It is easy to fall into local minimum and does not significantly improve the recognition performance of CNN.On the contrary,using the ROI segmented vessel intima and adventitia to pre-train CNN and fun-tuning the parameters of CNN by transfer learning can effectively improve the recognition performance of CNN,and the area under the curve(AUC)of receiver operating characteristic(ROC)increases from 0.862 to 0.900.Finally,first-order gray statistics,gray-level co-occurrence matrix,Laws texture energy and the other four types of texture features,a total of 85,were extracted from original 1,828 samples.And these features are used as the inputs to the support vector machines(SVM)to classify carotid artery plaque ultrasound images.Compared with the results of the CNN method,the experimental results show that CNN is obviously superior to the traditional SVM method based on feature extraction.