total of 105 lesions, 70 malignant and 35 benign, presenting as non-mass-like
enhancements were analyzed. Two radiologists gave the BI-RADS reading for the
morphological distribution and the internal enhancement pattern. For each case,
the 3D tumor mask was generated using FCM clustering algorithm with connective
labeling and hole filling. Three DCE parameters maps were generated from the
images, and PyRadiomics was applied to extract a total of 321 features for each
case. The diagnostic model was built using SVM with 10-fold cross-validation.
The accuracy of the radiomics model was 82%, higher compared to 72% built with the