Clinical Study Of Deep Neural Network Based On Ultrasonic RF Signal In Breast Tumor Diagnosis

ObjectiveBased on the clinicopathologic information of breast tumor and the corresponding Radio Frequency(RF)time series signal characteristic values,this study used machine learning to build a deep neural network model to evaluate the feasibility of the model for the diagnosis of breast tumor.And to verify that the results of applying the model to diagnose breast tumors outperformed the prototype ultrasound.MethodsA total of 300 patients with breast tumors confirmed by puncture or surgical pathology were collected from the Department of Breast Surgery,China-Japan Friendship Hospital,Jilin University from January 2021 to June 2022.According to the inclusion and exclusion criteria,262 eligible patients with breast tumors were screened.VINNO8 color ultrasound imaging system was used in all patients to collect ultrasonic images and ultrasonic RF time series signals.The collected patients were randomly divided into group A and Group B.Group A was the machine learning group,containing 160 patients,including 59 benign tumors and 101 malignant tumors;Group B was the clinical study group,containing 102 patients,including 38 benign tumors and 64 malignant tumors.In this study,fast Fourier transform(FFT)was used to convert the collected ultrasonic RF time series signals of patients in group A and B from time domain information to frequency domain information.After low-pass filtering,characteristic values in frequency domain were extracted and calculated,namely,maximum value,average value,variance and root mean square.The inverse Fourier transform is used to convert the filtered frequency-domain information into time-domain information,and the time-domain characteristic values,namely,kurtosis,skewness,crest factor,form factor,pulse factor and marginal factor,are extracted.The extracted characteristic parameters of group A tumor patients were used for machine learning,and the Levenberg-Marquardt method of Matlab neural network toolbox was used to construct the deep neural network.The mean square error and regression R value of the output result and target value of the neural network are used as the comprehensive evaluation of the prediction accuracy of the network.The trained deep neural network derived code and generated neural network model for clinical diagnosis of breast tumor patients.In order to verify the feasibility of the generated neural network model,the parameters of the tumor patient characteristics of group B were input into this model,and the output diagnostic results were subjected to Kappa concordance test with the pathological results.Subsequently,this diagnostic result was subjected to Kappa concordance test with the prototype ultrasound results,and the diagnostic test was validated on the basis of the concordance.Sensitivity represents the probability of correct detection of breast malignant tumor,specificity represents the probability of correct detection of breast benign tumor.The detection rate and diagnostic accuracy of benign and malignant breast tumor between this model and prototype ultrasound are compared and analyzed.ResultsThe machine training results of the time domain and frequency domain characteristic parameters of patients with breast tumor in group A showed that the mean square error between the output of training set,verification set and test set and the target were 0.0385,0.0582 and0.0646,respectively,and the regression R values were 0.914,0.859 and 0.842,respectively,indicating that the accuracy of network prediction was good.The extracted feature parameters are different.For patients with breast tumor in group B,the Kappa value of model output and pathological results was 0.793,which was in good agreement,indicating that this model could be used for the diagnosis of benign and malignant breast tumors in clinical practice.The Kappa value of model output and traditional ultrasonic results was 0.837,which was in good agreement.The results of the validation study showed that the sensitivity,specificity and accuracy of the model diagnosis were 90.6%,89.5% and 90.2%,respectively.The sensitivity,specificity and accuracy of ultrasonic diagnosis were 82.8%,81.6% and 82.4%,respectively,indicating that the model was superior to the prototype ultrasound in the diagnosis of breast tumors.Conclusions1.There are differences in the time domain and frequency domain characteristic values of ultrasonic RF time series signals of benign and malignant breast tumors;2.The deep neural network model based on breast ultrasound RF time series proposed in this study can be used for the diagnosis of benign and malignant breast tumors;3.The diagnostic results of the RF time series deep neural network model proposed in this study are superior to the original ultrasound,which can assist ultrasound to improve the diagnostic efficiency of breast tumors.