Research On The Construction Of Prediction Model Of Breast Malignant Lesions Based On Mammography

Objective: It is difficult to distinguish benign and malignant microcalcifications in the non-mass mammary gland,and to discuss the value of mammography imaging in the diagnosis of benign and malignant breast microcalcification lesions based on preoperative mammography and intraoperative pathological tissue mammography images,respectively,so as to discuss the value of mammography radiomics in the diagnosis of benign and malignant breast microcalcification lesions.Materials and methods: A retrospective collection of female patients with clear pathological results from August 2015 to July 2022 with non-mass breast microcalcifications as the main complaints,a total of 235 cases and 247 lesions in the preoperative group and 184 cases and 191 lesions in the intraoperative group,the clearest mammogram images of the microcalcified lesions were obtained and stored in DICOM format.All pathologies were divided into training and test sets in a 7:3 ratio according to random stratified sampling.Breast surgeons and radiologists work together to segment mammography images(ROI)on the DARWIN intelligent scientific research platform.The texture features and grayscale features of the wavelet processed image in the ROI region were extracted,and the optimal image features above the threshold were screened out by the variance selection method for model building.Four machine learning algorithms were used to establish the prediction model,the optimal parameters of grid search,the working characteristic curve(ROC)of each model subject,and the area under the ROC curve(AUC),sensitivity,specificity,F1-score and accuracy of each model were calculated,and the models with the best predictive performance in the preoperative group and intraoperative group were selected.Finally,the prediction performance of the intraoperative model established by the preoperative mammogram was evaluated as an independent verification set,and the prediction performance of the intraoperative model was compared with the prediction performance of the intraoperative model,and the performance difference was evaluated by the Delong test to verify the generalization ability of the intraoperative model to the preoperative mammogram image.Results: 1.A total of 247 non-mass microcalcification lesions from 235 female patients in the preoperative group were included in this study,with pathological results as the gold standard,158 cases(63.97%)were benign and 89(36.03%)were malignant;A total of 191 lesions were included in 184 female patients in the intraoperative group,including 121 cases(63.35%)and 71 cases(37.17%)of malignant and malignant,and the proportion of benign and malignant was basically in line with real-world statistics.2.For the preoperative group model,the AUCs in the training set based on K-NN,SVM,RF and GBDT were 0.77,0.81,0.95 and 0.97,and the ACCs were 0.70,0.70,0.90 and0.93,respectively;the AUCs in the test set were 0.58,0.61,0.64,0.65,and the ACCs were 0.72,0.68,0.73 and 0.69,respectively.For the intraoperative group model,the AUCs of the four classifiers in the training set were 0.81,0.84,0.94 and 0.98,and the ACCs were 0.73,0.80,0.89 and 0.93,respectively,while the AUCs in the test set were0.66,0.82,0.77 and 0.71,and the ACCs were 0.68,0.79,0.82 and 0.85,respectively.The overall efficiency of the intraoperative group model was significantly higher than that of the preoperative group model,and the model based on RF classifier was better than that of other classifiers.3.Among the texture features,the gray symbiotic matrix(GLCM)feature may be the most significant feature in constructing a classification model of benign and malignant breast microcalcification lesions.4.In the independent verification of the preoperative mammogram image dataset by the preoperative group model and the optimal model of the intraoperative group,the AUC was 0.87 and 0.91,the ACC was 0.82 and 0.88,respectively,and the Delong test Z=2.48,P=0.01,and the diagnostic performance of the intraoperative group model was better for the preoperative mammogram,and the difference was statistically significant.Conclusion: Based on preoperative mammogram imaging and intraoperative pathological tissue mammography imaging,four different machine learning algorithms were used to construct radiomics models,namely K-NN,SVM,RF and GBDT,and the model performance based on RF and GBDT algorithms was more stable,and RF was slightly better than GBDT,showing good benign and malignant prediction performance of breast microcalcification lesions.The model based on intraoperative pathological mammogram is significantly better than the model based on preoperative mammogram.Moreover,the intraoperative group model can maintain good predictive performance for preoperative mammogram images,and the generalization ability is strong.