Assessment of machine learning model performance to differentiate benign and malignant breast lesion: Finding best radiomic features on MDME MRI

Motivation: Lesion characteristics were investigated by MDME MRI-derived tissue-relaxometry, however, radiomics features of MDME generated images in predicting breast lesion malignancy has not been explored much Goal (s): The aim of this study is to find the best imaging features from MDME generated images in distinguishing malignant lesion and compare its performance with BI-RADS Approach: ML algorithms were explored and best-performing model using clinical-features, radiomic-features of four saturation-delay and two-echos scanned before and after contrast injection. Results: Test prediction based on BI-RADS achieved AUCs of 0. 67 in contrast best-performing stacking model achieved AUCs of 0. 82 using image radiomic-features of two-echoes of 2nd-saturation delay and clinical-features. Impact: Comparing with BIRADS, post contrast MDME derived radiomics-based machine learning shows promising potential in differentiating malignant breast lesion. Which may simplify of breast image scanning protocols and pulse-sequence-design for malignancy check.