Segmenting residual enhancing tumor and FLAIR hyperintensity in early postoperative glioblastoma MRI using AI, with EOR assessment and survival analysis

Among adult brain tumors, GBM is both relatively prevalent and highly fatal, for which surgery is part of standard of care. In this project, AI models are developed to segment residual CET and FLAIR hyperintensity in early post-op GBM mpMRI taken < 72 hours after surgery, with the segmented CET volume used to classify extent of resection. These can reduce workload of clinicians. Pretraining was done on MRIs from 882 cases taken from BraTS 2024 post-treatment glioma challenge dataset and models were finetuned using five-fold cross validation on early post-op mpMRI taken for 67 patients from the Oslo University Hospital BrainPower database. The ensemble of five models achieved a mean dice of 0.49 (SD: 0.34) for CET and 0.82 (SD: 0.16) for FLAIR hyperintensity on 20 hold out patients. Classification between maximal and submaximal resection achieved precision/recall of 1.00/0.92 for maximal and 0.88/1.00 for submaximal resection on the 20 hold out patients. On the external test set of 40 patients' MRI taken from the Río Hortega University Hospital GBM dataset, performance was 0.92/0.97 for maximal and an unreliable 0.50/0.25 for submaximal resection as the external data has severe class imbalance with only 4 submaximal resections. Segmentation performance on longitudinal MRI from the SAILOR dataset at OUH is explored and analyzed. Potential use of radiomic shape features extracted from CET and FLAIR hyperintensity noted on early post-op MRIs is explored for survival modeling using data of 80 patients. Kaplan meier curves show that patients with maximal resection tend to have longer overall survival than those with submaximal resection with a borderline log-rank p = 0.063. Comparative analysis using AICc, Akaike weights, likelihood ratio tests, and performance metrics across Cox PH and early death classification models reveals that actual residual CET volume outperforms surgical extent of resection in overall survival modeling in this cohort.