Research And Application Of Deep Learning In Carotid Ultrasound Image Diagnosis

Carotid ultrasound image diagnosis is an important basis evidence for the diagnosis and treatment of cardiovascular and cerebrovascular diseases,which is one of the important auxiliary means for doctors to diagnose and treat patients.It directly impacts the doctor’s treatment for patients and plays a vital role in the recovery and cure of patients.However,the current population base of China is too large and there are many problems need to be solved,such as the shortage of radiologist and the enormous pressure of outpatient visit.Moreover,the final diagnostic results will be influenced by the professional level and working time of radiologist.Therefore,computer-aided diagnos through the deep learning method will effectively improve the accuracy of diagnosis and deal with the staff shortages.Based on the characteristics of medical ultrasound image,this paper develops the research and application of automatic diagnosis of carotid plaque ultrasound image lesions on the basis of image processing,deep learning and other theories.The main research contents are as follows:(1)We analyze the characteristics of carotid ultrasound images such as histogram and edge gradient.For carotid ultrasound images with the problems of less data and imbalance,we adopt the Adding Gaussian noise,random brightness adjustment,random image blurring and other method to achieve the reasonable data augmentation.Experiments show that the data-enhanced image plays an important role in follow-up deep learning training models,so that the deep learning model has better convergence.(2)On account of the classification and location of complex carotid plaque target area in different directions of the same patient,a parallel branch plaque target detection algorithm based on convolutional neural network is proposed: For the two transverse and longitudinal sections of ultrasound images,the feature maps are extracted using the parallel branches of two pre-trained convolutional neural networks.Classification and localization are completed according to the features of the plaque area in the feature maps,and the carotid arteries in both transverse and longitudinal directions are effectively solved in view of problems for automated plaque diagnosis.In the network training process,We use transfer learning,batch regularization,residual neural networks and other method to make the model quickly converge and improve the training effect of the network.The final evaluation showed that the AP of the transverse and longitudinal plaques and blood vessels are above 0.84,and the F2-score of the transverse plaques and blood vessels reach 0.89 and 0.92 respectively.(3)The rectangular area of the carotid ultrasound image plaque is obtained according to the deep learning object detection algorithm,and the carotid plaque area was further segmented by example.The paper proposes an active contour automatic segmentation algorithm based on energy function for plaque detection on deep learning plaque detection.The evaluation results show that DSC is above 0.82 and HD below 6.2.Comparison with the doctor’s manual segmentation evaluation results shows that it can meet the application needs.The IMT thickness of carotid plaques was calculated and measured,and the degree of plaque risk was further evaluated through the study of carotid plaque risk level.