Fully automatic classification of breast lesions on multi-parameter MRI using a radiomics model with minimal number of stable, interpretable features.

Abstract

To build an automatic computer-aided diagnosis (CAD) pipeline based on multiparametric magnetic resonance imaging (mpMRI) and explore the role of different imaging features in the classification of breast cancer. A total of 222 histopathology-confirmed breast lesions, together with their BI-RADS scores, were included in the analysis. The cohort was randomly split into training (159) and test (63) cohorts, and another 50lesions were collected as an external cohort. An nnUNet-based lesionsegmentation model was trained to automatically segment lesion ROI, from which radiomics features were extracted for diffusion-weighted imaging (DWI), T2-weighted imaging (T2WI), and contrast-enhanced (DCE) pharmacokinetic parametric maps. Models based on combinations of sequences were built using support vector machine (SVM) and logistic regression (LR). Also, the performance of these sequence combinations and BI-RADS scores were compared. The Dice coefficient and AUC were calculated to evaluate the segmentation and classificationresults. Decision curve analysis (DCA) was used to assess clinical utility. The segmentation model achieved a Dice coefficient of 0.831 in thetest cohort. The radiomics model used only three features from diffusion coefficient (ADC) images, T2WI, and DCE-derived kinetic mapping, andachieved an AUC of 0.946 [0.883-0.990], AUC of 0.842 [0.6856-0.998] in the external cohort, which was higher than the BI-RADS score with an AUC of 0.872 [0.752-0.975]. The joint model using both radiomics score and BI-RADS score achieved the highest test AUC of 0.975 [0.935-1.000], with a sensitivity of 0.920 and a specificity of 0.923. Three radiomics features can be used to construct an automatic radiomics-based pipeline to improve the diagnosis of breast lesions and reduce unnecessary biopsies, especially when using jointly with BI-RADS scores.