GLISTRboost
Overview
GLISTRboost is a hybrid generative-discriminative brain tumor MRI segmentation algorithm that won the BraTS’15 challenge. It combines a generative probabilistic tumor growth model (GLISTR) with a discriminative gradient-boosting classifier to delineate three tumor sub-regions from multi-parametric MRI: enhancing tumor (ET), non-enhancing tumor core (NET), and peritumoral edema (ED). Input modalities are T1, T1-Gd (post-contrast), T2, and T2-FLAIR. GLISTRboost is used to generate initial automated segmentation labels that are then refined by expert annotators for large-scale imaging data releases. See also brats-challenge for the benchmark framework and captk for the software toolkit used alongside GLISTRboost for feature extraction.
Used by
- PMID:28872634 — GLISTRboost used to generate initial tumor sub-region labels (ET, NET, ED) for 243 pre-operative mMRI scans (TCGA-GBM n=135, TCGA-LGG n=108); automated labels were then manually revised by a board-certified neuroradiologist; median DICE on BraTS’15 GBM training set was 0.92 (whole tumor), 0.88 (tumor core), 0.88 (enhancing tumor); median Jaccard agreement between automated and final manually-revised labels was 0.96 (whole tumor), 0.87 (tumor core), 0.86 (enhancing tumor) PMID:28872634.
Notes
- Manual revision impacted the core/enhancing-tumor labels substantially more than the whole-tumor extent, reflecting the higher segmentation uncertainty in ET and NET boundaries PMID:28872634.
- The GLISTRboost-generated labels, after manual revision, became the reference labels used in the BraTS’17 challenge PMID:28872634.
- Ambiguous regions at the NET/ED boundary were left as automated labels without manual correction, introducing a known noise floor in fine-grained sub-region labels PMID:28872634.
Sources
This page was processed by crosslinker on 2026-05-04.