Imaging Signature of 1p/19q Co-deletion Status Derived via Machine Learning in Lower Grade Glioma

摘要

We present a new approach to quantify the co-deletion of chromosomal arms lp/19q status in lower grade glioma (LGG). Though the surgical biopsy followed by fluorescence in-situ hybridization test is the gold standard currently to identify mutational status for diagnosis and treatment planning, there are several imaging studies to predict the same. Our study aims to determine the lp/19q mutational status of LGG non-invasively by advanced pattern analysis using multi-parametric MRI. The publicly available dataset at TCIA was used. Tl-W and T2-W MRIs of a total 159 patients with grade-II and grade-III glioma, who had biopsy proven lp/19q status consisting either no deletion (n = 57) or co-deletion (n = 102), were used in our study. We quantified the imaging profile of these tumors by extracting diverse imaging features, including the tumor's spatial distribution pattern, volumetric, texture, and intensity distribution measures. We integrated these diverse features via support vector machines, to construct an imaging signature of 1p/19q, which was evaluated in independent discovery (n = 85) and validation (n = 74) cohorts, and compared with the lp/19q status obtained through fluorescence in-situ hybridization test. The classification accuracy on complete, discovery and replication cohorts was 86.16%, 88.24%, and 85.14%, respectively. The classification accuracy when the model developed on training cohort was applied on unseen replication set was 82.43%. Non-invasive prediction of 1p/19q status from MRIs would allow improved treatment planning for LGG patients without the need of surgical biopsies and would also help in potentially monitoring the dynamic mutation changes during the course of the treatment.