Prediction of genetically-evaluated tumour responses to chemotherapy from breast MRI using machine learning with model selection. Issue 1 (February 2021)

• Record Type: Journal Article
• Authors: Yuan, Taiguang; Jin, Ze; Tokuda, Yukiko; Naoi, Yasuto; Tomiyama, Noriyuki; Obi, Takashi; Suzuki, Kenji
• Is Part Of: Journal of physics. Volume 1780:Issue 1(2021)
• Publication Date: 2021-02
• Subjects: Physics -- Congresses; 530.5
• DOI: 10.1088/1742-6596/1780/1/012040
• Languages: English
• ISSNs: 1742-6588

Abstract

Abstract: Genetic tests can provide prognostic information in breast cancer for both diagnosis and treatment planning. However, the cost of a genetic test is still high. In this study, we developed a radiogenomics method to predict genetically-evaluated responses to chemotherapy for breast cancer using our machine-learning technology coupled with model selection. Our proposed method consists of feature extraction, model selection, and prediction by the selected model. In the feature extraction, 318 morphological and texture features were extracted from a tumour region. In the model selection module, there are two major components: (1) selection of imaging biomarkers based on our original sequential forward floating selection (SFFS) feature selection and (2) building of a support vector machine (SVM) classifier including kernel function selection and hyperparameter optimization. The optimized feature set, i.e. imaging biomarkers, coupled with an SVM classifier were chosen by maximizing the area under curve (AUC) of corresponding receiver-operating-characteristic curve (ROC). After the model selection, the optimized SVM classifier operated on the selected imaging biomarkers for prediction. We applied our proposed method to 118 breast MRI studies from 118 patients for predicting genetically-evaluated responses to chemotherapy for breast cancer that evaluated by the genetic test of IRSN-23. We achieved an AUC value of 0.96 using the optimized SVM classifier model coupled with 24 selected imaging biomarkers in predicting the results of IRSN-23 in a five-fold cross-validation procedure.