ML-Based Radiomics Analysis for Breast Cancer Classification in DCE-MRI

Abstract

Breast cancer is the most common malignancy that threatening women’s health. Although Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) for breast lesions characterization is widely used in the clinical practice, physician grading performance is still not optimal, showing a specificity of about 72%. In this work Radiomics was used to analyze a dataset acquired with two different protocols in order to train Machine-Learning algorithms for breast cancer classification. Original radiomic features were expanded considering Laplacian of Gaussian filtering and Wavelet Transform images to evaluate whether they can improve predictive performance. A Multi-Instant features selection involving the seven instants of the DCE-MRI sequence was proposed to select the set of most descriptive features. Features were harmonized using the ComBat algorithm to handle the multi-protocol dataset. Random Forest, XGBoost and Support Vector Machine algorithms were compared to find the best DCE-MRI instant for breast cancer classification: the pre-contrast and the third post-contrast instants resulted as the most informative items. Random Forest can be considered the optimal algorithm showing an Accuracy of 0.823, AUC-ROC of 0.877, Specificity of 0.882, Sensitivity of 0.764, PPV of 0.866, and NPV of 0.789 on the third post-contrast instant using an independent test set. Finally, Shapley values were used as Explainable AI algorithm to prove an high contribution of Original and Wavelet features in the final prediction.