Apparent diffusion coefficient map based radiomics model in differentiating benign from malignant entity in breast imaging-reporting and data system 4 breast lesions

Abstract

Objective To investigate the diagnostic values of radiomics model based on ADC map in differentiating benign from malignant lesions in suspicious breast findings with MRI breast imaging reporting and data system (BI-RADS) category 4. Methods Eighty eight patients (36 benign and 52 malignant) with diagnosis of MRI BI-RADS 4 in our hospital from December 2014 to December 2015 were retrospectively enrolled in this study. All the patients were proved by pathology and examined by sequence of T1WI, T2WI, DWI and dynamic contrast enhanced magnetic resonance imaging (DCE-MRI), and then were further sub-categorized into BI-RADS 4A, BI-RADS 4B, or BI-RADS 4C based on DCE-MRI sequence. Thus positive predictive value (PPV) of each sub-category were calculated and ROC were used to describe its efficiency in differential diagnosis. Radiomics features based on ADC map were generated automatically from Analysis-Kinetics (GE Healthcare). Sixty three of the 88 cases randomized-selected by computer were used to establish forecasting models and other 25 cases for validation. Kruskal-Wallis test and Spearman were introduced to reduce radiomics features that were highly correlated with others. Logistic linear regression (LLR) model was constructed based on the selected features by 'glm’ function in R software and then verified by 10-fold cross validation (repeat 10 times). ROC was curved to determine the diagnostic accuracy of the model. Results The PPV of BI-RADS 4A,4B,4C were 16.7% (2/12) ,59.6% (28/47) ,75.9% (22/29) respectively. Area under curve (AUC) of ROC was 0.650, sensitivity and specificity were 76.9% and 45.9% respectively. Three hundred and ninety six radiomics features were extracted automatically by software and 5 features (size zone variability, difference entropy, zone percentage, intensity variability and inverse difference moment) were left after redundancy reduction step. Cross-validation showed the accuracy of Logistic regression model was 80.0% (20/25), AUC was 0.790, sensitivity and specificity were 81.3% and 77.8% when cut-off was 0.45. Conclusion The radiomics model could provide important reference for differentiation between benign and malignant lesions in suspicious BI-RADS 4 findings. Key words: Breast neoplasms; Magnetic resonance imaging; Radiomics