Convolutional Neural Network Based Medical Ultrasound Image Quality Assessment

The quality of ultrasound medical images is an essential excellence evaluation index of ultrasonic imaging equipment and image processing algorithms.Though it is very important,there is no uniform standard of ultrasound image quality assessment(IQA)because of its high subjectivity.In this paper we try to use convolutional neural networks(CNNs)to settle our problem.Using CNN models to simulate the perception process of doctors on evaluating ultrasound images and achieve the goal of predicting quality of medical ultrasound images in a quantitative manner.Based on the analysis of the background,research significance and current situation of medical ultrasound IQA,this paper briefly introduces the applications of CNN in the field of optical IQA.And then we try to use CNN in medical ultrasound IQA.The main work includes the establishment of ultrasonic image database and the training of ultrasonic IQA model under three different CNNs.Firstly,we obtain the original ultrasound image database which contains 1063 images through the collection of original high-quality medical ultrasound images and artificial quality degradation of them.After doctors' evaluation,score processing,rejecting outliers and distribution balancing,478 ultrasound images with their subjective score labels are used as training and testing samples.On the basis of ultrasound image database,we use shallow CNN,deep CNN as well as adjusted Residual CNN to train objective IQA models,respectively.Transfer learning based on natural images is also used for model optimization.In the test,the linear correlation coefficient of subjective scores and model predictions is 0.832,and the Spearman's rank correlation coefficient of them is 0.797.The results are much better than the traditional objective evaluation methods such as PSNR and SSIM.The study of this paper shows that it is feasible and effective to learn the subjective quality evaluation of medical ultrasonic images by CNN,and it is expected to achieve practical applications of CNN evaluation models in the follow-up research.