Breast Tumor Genes Subtype Profiling Using MR Image Radiomic Features and Machine Learning Algorithms

Breast magnetic resonance imaging (MRI) Radiogenomics is an emerging area with the potential to influence clinical diagnosis and prognosis. Altogether, 106 MRI images collected from TCIA were studied to predict four Genes of (estrogen receptor (ER), progesterone receptor (PR), hormone receptor (HR), and human epidermal growth factor receptor-2 (HER2)) expression phenotypes from MR images. We extracted 104 radiomic features, including first-order statistical, morphological, and textural features, from pre-processed breast MR images using the Pyradiomics python library. In this work, a total of 5940 computational strategies composed of 33 feature selection methods, 12 machine learning models, and 15 resampling algorithms have been cross-combined and examined in our predictive model. The performance of different models was compared in terms of area under the curve-receiver operator characteristic (ROC-AUC) in addition to sensitivity and specificity parameters. Her2 was the best-predicted gene from MR images with a ROC-AUC of 0.87, while ER, HR, and PR were predicted with ROC-AUC factors of 0.69, 0.61, and 0.54, respectively. We demonstrated that radiomic features derived from MR images could be used for gene subtype profiling in breast cancer patients, specifically in predicting the Her2 type gene.