Automatic prediction of isocitrate dehydrogenase mutation status of low-grade gliomas using radiomics and domain knowledge inspired features in magnetic resonance imaging

Abstract

Aim: Most common and most deadly primary central nervous tumors, glial tumors harbor many heterogeneous clones of cells. Noninvasive determination of the genomic profiles of these tumors would have important implications regarding the classification, management, and prognostication of these tumors. Isocitrate dehydrogenase mutation is a key genomic signature that can downgrade the expected dismal course of these tumors. In this study we aimed to build a performant prediction model which can determine the Isocitrate Dehydrogenase (IDH) mutation status of glial tumors, using radiomics and leveraging automatic computation of domain knowledge-inspired features. Methods: Radiomics methods based on high throughput feature extraction and application of data science principles to these extracted features are promising tools for the noninvasive classification of lesions. Domain knowledge-inspired features besides radiomics features can contribute positively to the performance of the models. Some efforts particularly a joint approach to standardize the magnetic resonance imaging (MRI), reporting of glial tumors are mainstay for domain knowledge-inspired features. However, this requires active involvement and reporting of the radiologist which hampers automatization efforts. Additionally, this feature set evaluates a small subset of all possible signal and spatial-based computations. In this study, we combined domain knowledge-inspired features with radiomics features along with a multiparametric multihabitat comprehensive lesion description strategy. Results: Our best model which consisted of a combination of radiomics, and radiologist knowledge-inspired features reached a 0.93 f1 score (standard deviation (SD): 0.03), 0.93 accuracy (SD:0.03), and 0.98 area under curve (AUC), (SD:0.02). Conclusion: The multiparametric and multiregional approach employed in this study coupled with the integration of both radiomics and domain knowledge-inspired features resulted in a high-performance model emphasizing the contribution of each strategy to the outcome.