Purpose: True proliferation of gliomas can only be measured by biopsy and is prone to sampling error due to tumor heterogeneity. We estimate tumor proliferation using MR imaging then demonstrate the ability for predictions to inform prognosis and guide surgical treatment.
Methods: 52 image-guided(sub-millimeter accuracy) biopsy samples from 23 patients were used to train a random forest model to estimate proliferative index(%Ki-67 expression). Model inputs were routine MRI sequences: T1w (pre- and post-contrast), T2w, and FLAIR. This model was applied to a glioma patient cohort from our institution containing 207 grade II, 246 grade III, and 728 grade IV tumors. MR images for these patients were co-registered and normalized using normal tissue intensities. The random forest was applied voxel-wise through the automatically segmented tumor volume to estimate the pointwise proliferative index. We then calculated the maximum estimated proliferative index with at least 0.1cc volume in each tumor. For 811 cases with postoperative imaging available we also calculated the maximum postoperative proliferative index within unresected tumor. Maximum estimated proliferation was correlated with survival using univariate and multivariate(including mental status, age, grade) Cox models.
Results: The root-mean-square-error for predicting Ki-67 with the random forest was 5.4%. A threshold of 11.6% on maximum preoperative proliferative index showed strong survival differences between patients with high and low proliferation tumors: univariate hazard ratio of 2.18(log-rank p<<0.05), multivariate hazard ratio 1.38(p<1E-4). Furthermore, reduction in highly proliferative disease postoperatively was associated with improved survival. Patients with high maximum postoperative proliferation had significantly worse survival(p<0.05) than patients where highly proliferative disease was removed.
Conclusion: MR imaging non-invasively predicts pointwise proliferative activity. Estimated proliferation also predicts survival in glioma patients independently of clinical factors and tumor grade. Changes after surgery are prognostic, supporting the use of estimated proliferation to both guide and assess surgical intervention.
Funding Support, Disclosures, and Conflict of Interest: Evan Gates is supported by a training fellowship from the Gulf Coast Consortia, on the NLM Training Program in Biomedical Informatics & Data Science (T15LM007093).
IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)