Research On Breast Ultrasound Image Segmentaion Based On Residual U-shaped Convolution Neural Network

Breast cancer,which is one of the most common malignant tumors,severely impairs female health.The key of the treatment is receiving diagnosis at the early stage.Ultrasonic imaging examination is an important non-intrusive diagnosing method,relying on the doctor’s interpretation of ultrasound images,the time required is long and the workload is heavy,and it barely meets the needs of rapid and mass clinical diagnosis.To raise the diagnosing efficiency of ultrasonic imaging examination and reduce the burden of doctors,Computer-Aided Diagnosis(CAD)basing on breast ultrasound images comes to rescue.Image segmentation is used to extract tumor regions from normal tissue areas such that the interference from normal tissues around tumor region can be minimized.The performance of image segmentation affects accuracy and reliability of diagnosing results in a straightforward way.It is an indispensable part of Computer-Aided Diagnosis(CAD)system.With the flourish development of deep learning in recent years,image segmentation using deep learning has better performance than traditional segmentation methods in terms of automatic feature extraction and segmenting precision.U-Net,which is a kind of convolutional neural network architecture,has been widely applied in medical image segmentaion.However,due to the fact that breast ultrasound images have features such as high noise,low contrast and weak boundary,it is very difficult to obtain desired tumor segmentation results by simply using U-Net.To improve the precision of segmentation,more layers are added to U-Net to make it deeper such that abilities of feature extraction and classification are enhenced in this work.Meanwhile,gradient vanishing problem that will typically appear when network goes deeper has been solved.The contributions of this dissertation are:1.An improved U-Net architecture,called ResU-Net,where residual unit from ResNet has been utilized,has been proposed.It has been adapted to the real application scenarios of image segmentation for breast tumors.ResU-Net is deeper than U-Net,and it greatly improves feature learning ability comparing to U-Net.Benefiting from ResNet unit,gradient vanishing problem has been solved,which makes training deeper networks much easier than before.Accordingly,segmentation precision has been improved significantly.2.A new data augmentation method using moving least squares transformation has been implemented.The difficulty of getting a large set of ultrasound images obtained from real application scenarios as well as the loss of diagnosing information when making use of traditional data augmentation methods have been considered.The proposed data augmentation method not only enlarges dataset scale but also ensures the integrity of important diagnosing information.3.203 breast ultrasound images which have the tumor areas tagged by doctors from Zhongshan University Affiliated Tumor Hospital are used as data source to measure the segmentation performance of ResU-Net.The experimental results show that ResU-Net is capable of segmenting ultrasound images into tumor regions and normal regions with great precision.The Similarity coefficient(Dice)of segmentation from testing dataset is 0.9568 and Intersection-over-Union(IOU)is 0.9173,which are better than segmentation models built with traditional convolutional neural networks.This indicates ResU-Net has a greate chance to be successfully applied to real diagnosing scenarios.