Room: AAPM ePoster Library
Establish the prognostic significance of MRI-based proliferative activity predictions in gliomas and evaluate the potential to improve survival by guiding surgical resection.
We applied a previously developed random forest model that was trained to predict the proliferative index (% Ki-67 expression) in spatially specific (<1 mm) biopsy samples using routine MRI sequences: T1w (pre and post-contrast), T2w, and FLAIR. Using these four images, the root-mean-square error was 5.4% for predicting the Ki-67 expression.
This model was applied to two independent patient cohorts with known outcomes: 1) 140 high-grade glioma cases publicly available through the 2018 BraTS challenge and 2) 68 previously untreated glioma patients that underwent surgical resection at our institution. Preoperative images were co-registered and normalized using reference tissue intensities and the random forest was applied voxel-wise through the visible tumor volume. For cohort 2, the same image processing was also performed on postoperative images acquired within two days of surgery. The maximum predicted Ki-67 expression in each tumor was correlated with survival using a Cox model and log-rank test.
The preoperative maximum Ki-67 predicted survival in both cohorts. Thresholds of 28.2% and 24.75% on the maximum estimated Ki-67 expression optimally stratified the cases with hazard ratios (HR) of 1.66 and 3.32 respectively (both p<0.05). Furthermore, cohort 2 demonstrated the potential to guide surgery: A high preoperative Ki-67 level combined with a low postoperative Ki-67 level (meaning the highly proliferative disease was removed) was favorably prognostic compared to high Ki-67 both pre- and post-op (HR=0.329, p<0.05).
Estimates of proliferative activity in gliomas based on brain MRI are predictive of survival. Our results suggest surgical removal of highly proliferative regions also improves survival. This supports using model predictions to guide surgical intervention. Future work will evaluate the potential benefit in the context of other clinical factors.
MRI, Brain, Quantitative Imaging
IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning