Ultrasonic Image Diagnosis Of Breast Tumors Based On The Fusion Of Depth And Shallow Features

Breast cancer is one of the common cancers in human beings.Early detection and diagnosis of breast cancer is a significant subject of medical research.With the development of convolutional neural networks,automatic end-to-end feature extraction has provided a great help for breast tumor ultrasound image diagnosis,substantially improving the objectivity and repeatability of clinical diagnosis.However,current auxiliary diagnosis techniques based on image algorithms still have been facing problems such as poor feature extraction and low recognition rate,due to complexity of the features of ultrasound images.In this paper,an in-depth study on the localization of breast tumor lesions and the fusion of deep and superficial features is carried out,on the basis of synthesizing relevant literatures at home and abroad.The main research contents are as follows:(1)The automatic localization algorithm for breast tumors was studied.In view of the current clinical problems of relying on physicians to manually select regions of interest and low tumor screening efficiency,a breast tumor ultrasound image dataset with segmentation markers was collected and produced in this chapter.Then,principles of several mainstream target detection algorithms under the framework of lesion area localization process were described,and experimental analysis and comparison were designed.Finally,combining quantitative analysis and qualitative analysis of each index,YOLOv3 model can accurately obtain suspicious lesions and realize automatic localization of breast tumors.(2)Deep feature extraction based on residual network and adaptive spatial feature fusion and breast tumor image classification algorithm was studied.Firstly,a multi-channel fusion preprocessing method was proposed to solve the problems of speckle noise and echo interference in the localized breast tumor ultrasound images.Canny operator description image and histogram description image containing various details were fused at the channel level to realize RGB color conversion and sharpness enhancement of image.Then,a two-stage tumor classification network with deep feature fusion was designed.The first stage based on improved residual network was used as the feature extraction backbone network,and the second stage adaptive spatial feature fusion algorithm was introduced to fuse low-level,middle-level and high-level feature information extracted by backbone network.Finally,Softmax layer was added to obtain the classification results.(3)A classification method based on feature fusion of depth feature and shallow local binary pattern was proposed.Aiming at problems of poor feature extraction and low recognition rate of breast tumors in ultrasound images,firstly an improved circular local binary pattern operator was used to encode complex texture information and extract texture features.Then,on the basis of depth feature extraction based on residual network and adaptive space feature fusion,depth feature and shallow texture feature were combined to complete classification by SVM.Finally,the proposed method was trained and verified on labeled breast ultrasound tumor images,of which 2288 were benign and 2086 were malignant.The experimental results could reach 0.9691 accuracy and 0.9910 AUC value.The nonlinear mapping algorithm was used to visualize the classification process and further determine the classification level of deep and shallow layer fusion network.