Research On Key Technologies Of Ultrasonic Breast Tumor Aided Diagnosis

Breast cancer is one of the most morbidity and most harmful malignant diseases for women in the world.Early screening and early treatment are effective methods to reduce mortality.Ultrasonography has become an important method for early breast cancer screening due to its advantages of non-radiation,non-invasiveness,low price,and portability.In traditional ultrasound examinations,doctors analyze the features of breast tumors in ultrasound images based on their clinical practice experience and match them with BI-RADS(Breast Imaging Reporting And Data System)standards to achieve clinical diagnosis of breast tumors.This is a subjective judgment process.Different doctors may have completely different diagnosis results for the same lesion due to differences in operating techniques,clinical experience,knowledge level,and fatigue levels.Therefore,how to use the computer-aided diagnosis system to improve the consistency of doctor’s disease interpretation and reduce the rate of misdiagnosis and missed diagnosis has important research value and clinical significance in the early screening of breast cancer.In view of this,this paper conducts research from three aspects of breast tumor of classification of benign and malignant,BI-RADS classification detection and calcification point detection,and proposes some improved algorithms to assist doctors in accurately diagnosing ultrasound breast tumors.The specific work content is as follows:(1)Aiming at the shortcomings of traditional quantitative feature description of breast tumors such as inaccurate morphology,a new morphological quantitative feature extraction method is proposed.First,the breast tumor area is segmented from the original ultrasound breast tumor image to obtain the breast tumor edge.Secondly,an ellipse or circle is adopted to fit the outline of the breast tumor and subtract the edge and the ellipse or circle to obtain the shape histogram of the ultrasonic breast tumor.Finally,based on the shape histogram,three new morphological quantitative features describing the breast tumor are designed: maximum curvature sum,maximum curvature peak sum,and maximum curvature standard deviation sum.The SVM(Support vector machine)classifier was used to perform experiments on 71 ultrasound breast tumor malignant pictures and 121 benign pictures,and the new morphological quantitative features were obtained to distinguish benign and malignant breast tumors with an accuracy rate of 82.69%,and traditional morphological quantitative features were 73.08%.The experimental results show that the new morphological quantitative features have higher classification accuracy than the traditional morphological quantitative features.(2)In view of the extremely unbalanced number of original data samples of BI-RADS grading images of ultrasound breast tumors and the indistinct discrimination of tumor features of different grades,the CH-ResNeSt50 network was designed to improve the accuracy of BI-RADS classification of ultrasound breast tumors.The network is improved on the basis of the ResNeSt50(Residual Split-Attention Network 50)network,by adding a cross-layer connection branch(CLC)module and a multi-scale hybrid dilated convolution(HDC)module to improve the differentiation of features of different grades of breast tumors.In the experiment on 1543 ultrasonic breast tumor pictures(774 pictures of grade 3,207 pictures of grade 4A,178 pictures of grade 4B,150 pictures of grade 4C,and 234 pictures of grade 5),the recognition accuracy of CH-ResNeSt50 was 81.18%.Compared to the original network ResNeSt50,CH-ResNeSt50 increased by 5.3%.The experimental results show that,compared with ResNeSt50,the final optimized network model CH-ResNeSt50 can achieve more accurate classification for the BI-RADS classification of ultrasonic breast tumors.(3)In the early screening of breast cancer,the accurate detection of ultrasound breast tumor calcification points has important clinical research value for assisting doctors in accurate diagnosis.In response to this problem,this paper proposes a method for detecting calcification points of ultrasound breast tumors based on multi-scale superpixel segmentation.First,a multi-scale superpixel segmentation algorithm is used to divide tumor regions of different sizes,and weak calcification points are detected according to the uneven gray-scale distribution and texture contrast between the regions to ensure that the boundaries of the detected ultrasonic breast tumor calcification points are as far as possible close to the edge of the real target.Secondly,single-scale superpixel segmentation is performed based on the original image,and strong calcification points are detected by comparing the difference in gray value and the calcification gray distance feature to ensure that the detected calcification points are accurate and reliable.Finally,by combining strong and weak calcification points,the final targets are accurate and the edges are tight.The detection algorithm proposed in this paper can effectively detect calcification points in breast ultrasound images and the accuracy of detecting calcification points in benign and malignant cases can reach81.8%.