Predicting the pathological status of mammographic microcalcifications through a radiomics approach

Objective The study aimed to develop a machine learning (ML)-coupled interpretable radiomics signature to predict the pathological status of non-palpable suspicious breast microcalcifications (MCs).

Methods We enrolled 463 digital mammographical view images from 260 consecutive patients detected with non-palpable MCs and BI-RADS scored at 4 (training cohort, n = 428; independent testing cohort, n = 35) in the First Affiliated Hospital of Nanjing Medical University between September 2010 and January 2019. Subsequently, 837 textures and 9 shape features were subsequently extracted from each view and finally selected by an XGBoost-embedded recursive feature elimination technique (RFE), followed by four machine learning-based classifiers to build the radiomics signature.

Results Ten radiomic features constituted a malignancy-related signature for breast MCs as logistic regression (LR) and support vector machine (SVM) yielded better positive predictive value (PPV)/sensitivity (SE), 0.904 (95% CI, 0.865–0.949)/0.946 (95% CI, 0.929–0.977) and 0.891 (95% CI, 0.822–0.939)/0.939 (95% CI, 0.907–0.973) respectively, outperforming their negative predictive value (NPV)/specificity (SP) from 10-fold cross-validation (10FCV) of the training cohort. The optimal prognostic model was obtained by SVM with an area under the curve (AUC) of 0.906 (95% CI, 0.834–0.969) and accuracy (ACC) 0.787 (95% CI, 0.680–0.855) from 10FCV against AUC 0.810 (95% CI, 0.760–0.960) and ACC 0.800 from the testing cohort.

Conclusion The proposed radiomics signature dependens on a set of ML-based advanced computational algorithms and is expected to identify pathologically cancerous cases from mammographically undecipherable MCs and thus offer prospective clinical diagnostic guidance.