The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification
Paper • 2107.02314 • Published
DEEP AG (Deep Adult Glioma)
DEEP AG is a two-stage 3D segmentation model for adult diffuse glioma on pre-operative brain MRI, released by Deep Autonomy. It was trained and evaluated on the BraTS 2023 Adult Glioma (BraTS-GLI) dataset.
Input is the four standard MRI sequences (T1, T1 post-contrast, T2, T2-FLAIR). Output is the three BraTS tumor regions: whole tumor, tumor core, and enhancing tumor.
Models
The release contains two model files.
| File | Role | Parameters | Size |
|---|---|---|---|
stage_one.pt |
Stage One, full-volume segmentation | ~47M | 189 MB |
stage_two.pt |
Stage Two, cropped refinement | 6.8M | 27 MB |
Stage One segments the whole MRI volume. Stage Two takes the region around the predicted tumor and refines the boundary. Together the two models total roughly 54M parameters and 216 MB.
Results
Official BraTS 2023 validation leaderboard
Scored by the BraTS 2023 Adult Glioma Synapse portal on the official validation set (219 cases, hidden ground truth), using the official lesion-wise metric.
| Region | Lesion-wise Dice | Lesion-wise HD95 |
|---|---|---|
| Whole tumor | 0.897 | 14.24 |
| Tumor core | 0.853 | 15.42 |
| Enhancing tumor | 0.807 | 30.75 |
| Mean | 0.852 | 20.14 |
Rank: approximately 301 of 3368 submissions (top 9%), and 22 of 174 participating teams, on the continuous-evaluation leaderboard.
This leaderboard result is the full system — the two-stage cascade released here plus an enhancing-tumor specialist ensemble. The two checkpoints in this repository are the two-stage base of that system; their standalone scores on our internal held-out split are below.
Internal held-out split (released checkpoints)
Lesion-wise Dice on a 125-case held-out split of the BraTS 2023 Adult Glioma training set, scored with the official BraTS 2023 lesion-wise metric.
| Region | Stage One | Stage One + Stage Two |
|---|---|---|
| Whole tumor | 0.854 | 0.874 |
| Tumor core | 0.727 | 0.737 |
| Enhancing tumor | 0.623 | 0.634 |
| Mean | 0.735 | 0.749 |
Files
Both checkpoints are plain PyTorch state dictionaries. Stage One embeds its model configuration so the architecture rebuilds on load.
Data and required citations
The models were trained and evaluated on the BraTS 2023 Adult Glioma (BraTS-GLI) challenge, distributed through Synapse (syn51156910). The Adult Glioma track uses the RSNA-ASNR-MICCAI cohort, so its data descriptor (reference 3 below) predates the 2023 challenge. Use of this model and any derived work must cite the following, as required by the BraTS data usage agreement:
B. H. Menze, A. Jakab, S. Bauer, et al. "The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS)." IEEE Transactions on Medical Imaging 34(10):1993-2024, 2015. doi:10.1109/TMI.2014.2377694
S. Bakas, H. Akbari, A. Sotiras, et al. "Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features." Scientific Data 4:170117, 2017. doi:10.1038/sdata.2017.117
U. Baid, S. Ghodasara, S. Mohan, et al. "The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification." arXiv:2107.02314, 2021. (Data descriptor for the Adult Glioma cohort reused by the BraTS 2023 Adult Glioma challenge.)
S. Bakas, H. Akbari, A. Sotiras, et al. "Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-GBM collection." The Cancer Imaging Archive, 2017. doi:10.7937/K9/TCIA.2017.KLXWJJ1Q
S. Bakas, H. Akbari, A. Sotiras, et al. "Segmentation Labels and Radiomic Features for the Pre-operative Scans of the TCGA-LGG collection." The Cancer Imaging Archive, 2017. doi:10.7937/K9/TCIA.2017.GJQ7R0EF
Challenge page: https://www.synapse.org/Synapse:syn51156910
Intended use
Research use only. This model is not a medical device and is not for clinical or diagnostic use.