Research On Key Technologies Of Computer-Aided Diagnosis For Medical Ultrasound Images Based On Deep Learning

Ultrasound images are widely used in clinical diagnosis regarding the advantages of the non-ionizing radiation,low cost,non-invasive and comfortable operation,but the diagnosis results depend largely on the doctor’s ability to interpret ultrasound images.Making a diagnosis based on ultrasound images is labor-intensive,so an accurate,reliable and efficient computer-aided diagnosis system can be used to assist doctors in diagnosis.The development of deep learning in the field of medical image can be applied for ultrasound image computer-aided diagnosis system.Computer-aided diagnosis system of medical ultrasound image based on deep learning is composed of three successive stages:image preprocessing,image segmentation and image classification.However,the research of medical ultrasound computer-aided diagnosis system faces great challenges due to the problems of low contrast,low signal-to-noise ratio and small image datasets.This thesis takes medical B-mode ultrasound image as the breakthrough point and focuses on the three key technologies of ultrasound image denoising,segmentation and classification in medical ultrasound image computer-aided diagnosis system.The main contents and innovations are as follows:1.Ultrasound images denoising based on generative adversarial networkThe inherent speckle noise of ultrasound image adversely affects the subsequent image analysis.To solve the problem of speckle denoising in the ultrasound images,this thesis proposes a denoising algorithm based on generative adversarial network,which includes a residual dense connectivity module for extracting image feature information,and the weighted joint loss can be used to optimize the denoising performance during the training process.Under the experiments of synthetic speckle image denoising at three noise levels,the peak signal-to-noise ratio and structure similarity of the proposed algorithm are improved by 2.66% and 1.68%,respectively,compared with the other eight denoising algorithms.And the experiments of speckle denoising were carried out on three types of B-mode ultrasound images(lymph node,fetal head,brachial plexus).The experimental results show that the proposed algorithm achieves the optimal denoising effect in both synthetic speckle noise image and real B-mode ultrasound image.2.Ultrasound image segmentation algorithm based on multi-channel atrous convolutionThe low resolution and low signal-to-noise ratio of ultrasound images can seriously affect the segmentation accuracy of ultrasound images,this thesis proposes an ultrasound image segmentation algorithm based on multi-channel atrous convolution,in which,multi-channel discrete large convolutional kernel is used to extract image feature information,while atrous convolution and pyramid pooling layer can be applied to fuse global and local feature information,so that solving the problem of segmentation target size differentiation.This thesis conducts image segmentation experiments on three types of B-mode ultrasound images(lymph node,fetal head,brachial plexus)and compares the Dice coefficient improved by 5.68% on average with the other four segmentation algorithms,thus achieving better segmentation accuracy.3.Classification of ultrasound image algorithm based on data augmentationThe number of datasets has a very important impact on the performance of deep learning algorithms,while the small image datasets is a major problem that deep learning algorithms has encountered in the field of medical images.To solve this problem,this thesis proposes a classification algorithm of ultrasound image based on image data augmentation.Firstly,conditional generative adversarial network is used to augment the ultrasound image segmentation labels.Secondly,based on image transformation algorithm,ultrasound images segmentation labels are converted into ultrasound images to change the traditional data augmentation algorithm.Compared with the traditional data augmentation algorithm based on affine transformation,the proposed data augmentation algorithm has an average improvement of 12.95% on accuracy,precision,recall and F1-score four evaluation metrics value of B-mode lymph node and breast ultrasound image benign and malignant tumor classification.This thesis puts forward some achievements in the research of the three key technologies of ultrasound image denoising,segmentation and classification,which provides support for the subsequent application of ultrasound image computer-aided diagnosis system.