NEIM-03. FEASIBILITY OF AUTOMATED ASSESSMENT OF PROGRESSIVE ENHANCEMENT ON MRI IN PATIENTS WITH NEWLY DIAGNOSED HIGH-GRADE GLIOMA USING A FEATURE-BASED ALGORITHM

Abstract

BACKGROUND Approximately 50% of patients with newly diagnosed high-grade glioma (HGG) develop progressive enhancement between their post-operative MRI scan and 12 weeks after radiation and temozolomide. Inter-reader variability on the assessment of progressive enhancement in this patient group is a significant barrier in designing multi-center biomarker trials to distinguish true progression from pseudoprogression. Although enhancement segmentation algorithms have become more widely available, more automated and reproducible techniques to identify patients who develop progressive enhancement are needed to facilitate acquisition of non-standard of care biomarkers when this occurs. We explored the feasibility of using a feature-based algorithm in tandem with freely available / open source automated segmentation algorithms to identify this subset of patients. METHODS An automated algorithm using subtraction of registered segmentations to detect new areas of localized thickness of enhancement was developed. Criteria for feasibility (50% within 95% CI of percent patients identified, and sensitivity of >85% of patients assessed as progressed [P+] identified) were determined prospectively. The algorithm was implemented across five different automated enhancement segmentation techniques, then evaluated using a retrospective dataset of 73 patients with newly diagnosed HGG (age 50.8±13.2 years, 37 men, 36 women, 50 GBM, 23 Grade III). Standardized post-baseline brain tumor imaging protocol MR acquisitions were obtained on 1.5T and 3T scanners (GE and Siemens). On chart review, 53% of patients were assessed by neuroradiologists and/or neuro-oncologists as P+ (progression vs. pseudoprogression). RESULTS 50% was within the 95% CI of percent of patients identified for all five segmentation algorithms. Sensitivity was over 85% for three segmentation algorithms, with the MIC-DKFZ algorithm having highest sensitivity of 92%. For this algorithm, specificity was 77%, PPV was 81% and NPV was 90%. CONCLUSION A feature-based algorithm in tandem with open source segmentation algorithms showed preliminary feasibility for automated identification of patients with progressive enhancement.