Key Technology Research Of Breast Tumor Grade Diagnosis Based On Ultrasound Radio-Frequency Signals

Malignant breast tumors are one of the tumors with the highest incidence in the female population in China today.Ultrasound is the most important method for distinguishing benign and malignant breast tumors.Based on clinical practice experience,sonographers analyze the tumor-related features in breast ultrasound images,and match them with the BI-RADS(Breast Imaging Reporting and Data System)standard,so as to realize the classification of benign and malignant breast lesions.It can be seen that the current manual reading method is a subjective judgment process,and the diagnosis result is limited by the doctor’s clinical experience,knowledge level,and fatigue level.The ultrasound radio frequency(RF)signal is the original echo signal received by the ultrasound probe.It has not been image-processed and contains the most comprehensive and true human tissue information.The current ultrasound examinations only use the amplitude information in the radio frequency signal.The very important information in the signal,such as frequency spectrum,phase,time variation,etc.,is ignored and lost during the imaging process.This article starts with the ultrasound radio frequency(RF)signal and studies different optimization algorithms for ultrasound breast tumor grading under different categories.The main research work done in this paper is as follows:(1)Study the segmentation of the tumor area from the breast ultrasound image to obtain the region of interest,and propose an automatic segmentation algorithm AE-Unet(Attentionenhancing Unet)based on the deep neural network for attention-enhancing U-shaped network.By introducing the attention mechanism,different weights are assigned to each attention mechanism to suppress the background area and highlight the tumor area.In addition,improve the network training method,use the overall loss function to train the overall network to make the network basically stable,and then use the partial loss function,the attention mechanism and the backbone network in rotation to improve the accuracy of the network parameters.Experiments show that the accuracy of this method reaches 93.85%,which can effectively improve the efficiency of tumor segmentation.(2)A multi-feature extraction algorithm for breast tumor grading that simulates the visual perception of breast tumors by sonographers is proposed.The algorithm uses BI-RADS as the grading basis,and extracts 16 feature parameters that are important for grading from the aspects of morphology,texture,and echo.In order to prevent over-fitting,Spearman‘s rank correlation coefficient is used for feature screening,and 4 other parameters: ’Ene’,’Cor’,’En’and ’Msd’.Machine learning of the remaining feature parameters,sorted by importance,analyze the relationship between feature parameters and tumor benign and malignant grades.Experiments show that the accuracy,sensitivity,and specificity of the feature parameters combined with the random forest model after adjusting the parameters for breast tumor classification and diagnosis reached 91.24%,91.62% and 90.15%,respectively.(3)In order to solve the limitation of tumor grading in the time domain,a breast tumor grading algorithm based on Shearlet transform is proposed in the frequency domain.The multi-scale and multi-directional features of different levels are extracted by the Shearlet transform multi-scale geometric method.Aiming at the directionality and sparsity of the Shearlet transform,the multi-scale directional binary pattern(MDBP)is used to reduce the dimensionality of its features.Under the precondition of not losing the existing feature information,the recognition speed is raised and the recognition efficiency is improved.In addition,a cascade binary tree SVM(CBT-SVM)classifier is constructed for the imbalance of different grades of breast RF signal data,which not only overcomes the problem of uneven sample distribution,but also conforms to clinicians’ reading habits.Experiments show that the accuracy of this algorithm in ultrasonic breast grading detection reaches 89.29%,and the positive and negative detection rates are 97% and 98.3%,respectively.(4)A breast tumor grading diagnosis algorithm based on the deep features of BIRADSNet is proposed.In the model establishment stage,in order to extract the superficial and deep features of the ultrasound image to the greatest extent,pruning and improving on the basis of the PANet model,a multi-scale breast tumor grading model with dual attention mechanism is proposed.In the feature extraction stage,Res Net 34 is used as a skeleton network and an attention machine(Convolutional Block Attention Module,CBAM)is added;in the feature fusion stage,an attention mechanism is added to further improve the accuracy of target detection and classification.In the model training stage,the original 928 images were dataenhanced and expanded to 5013 images.Experiments show that the accuracy and specificity of the method reach 92.19% and 95.76%,indicating that the BIRADS-Net model has a high accuracy rate in detecting both positive and negative results.