Background: Gliomas are the most common type of malignant brain tumor. Clinical outcomes depend on many factors including tumor molecular characteristics. Mutation of the isocitrate dehydrogenase (IDH) gene confers significant benefits in terms of survival and quality of life. Preoperative determination of IDH genotype can facilitate surgical planning, allow for novel clinical trial designs, and assist clinical counseling surrounding the individual patient's disease.
Methods: In this study, we aimed to evaluate a novel approach for non-invasively predicting IDH status from conventional MRI via connectomics, a whole-brain network-based technique. We retrospectively extracted 93 connectome features from the preoperative, T1-weighted MRI data of 234 adult patients (148 IDH mutated) and evaluated the performance of four common machine learning models to predict IDH genotype.
Results: Area under the curve (AUC) of the receiver operator characteristic were 0.76 to 0.94 with random forest (RF) showing significantly higher performance (p < 0.01) than other algorithms. Feature selection schemes and the addition of age and tumor location did not change RF performance.
Conclusions: Our findings suggest that connectomics is a feasible approach for preoperatively predicting IDH genotype in patients with gliomas. Our results support prior evidence that RF is an ideal machine learning method for this area of research. Additionally, connectomics provides unique insights regarding potential mechanisms of tumor genotype on large-scale brain network organization.
Keywords: IDH; MRI; connectomics; glioma; machine learning.