glioma-segresnet

3D SegResNet (MONAI, ~4.7 M parameters) for the segmentation of post-treatment glioma on four MRI sequences (T1, T1c, T2, FLAIR) into enhancing tumour (ET), non-enhancing tumour core (NETC), surrounding non-enhancing FLAIR hyperintensity (SNFH) and resection cavity (RC), used for longitudinal volumetry. This repository holds the weights and the example scans of the demo Space; the code lives in the GitHub repository the Space is deployed from.

Research use only. Not a medical device. Not for clinical decisions.

Files

  • best.pt: PyTorch checkpoint (loaded with gliomaseg.inference.load_checkpoint), SHA-256 78080d1a0c5057d7bc154cdd6e6bdf8ef8f2823a3d9d777625abd74bcbd360a1.
  • summary.json: examples, per-scan volumes and Dice, selection rule, provenance (git commit, config hash).
  • examples/<patient>/<timepoint>.npz: preprocessed example scans (uint8 images, prediction, reference).
  • NOTICE.md: data attribution and disclaimer.

Data

Trained on MU-Glioma-Post (TCIA, CC BY 4.0, DOI 10.7937/7k9k-3c83): skull-stripped, co-registered T1/T1c/T2/FLAIR at 1 mm with expert segmentations. Splits are patient-level; the examples are from the internal test split. MU-Glioma-Post: The University of Missouri Post-treatment Glioma MRI dataset, The Cancer Imaging Archive (TCIA), licence CC BY 4.0, doi:10.7937/7k9k-3c83.

Method

Preprocessing: RAS orientation, resampling to 1 mm if needed, brain crop, per-sequence percentile clipping and in-brain z-score. Training: Dice + cross-entropy on 96^3 patches. Inference: sliding window (96^3, 50 % overlap, Gaussian blending). Post-processing: small-component removal per class {'ET': 50, 'NETC': 10, 'SNFH': 200, 'RC': 100}. SegResNet: init_filters 16, blocks_down [1, 2, 2, 4], blocks_up [1, 1, 1].

Results (internal test split, mean Dice)

ET NETC SNFH RC TC WT Mean
0.787 0.682 0.864 0.586 0.755 0.867 0.757

TC = ET + NETC, WT = ET + NETC + SNFH.

Example selection

Among the test-split patients with >= 3 scans, the 3 patients whose patient-mean Dice (mean over scans of the mean region Dice ET/NETC/SNFH/RC/TC/WT, from the test per-case CSV) is closest to the median of the patient means. No visual inspection or manual choice.

Intended use and limitations

Intended for research and for model-assisted annotation studies and demonstrations. Limitations: a single data set from one institution; expects skull-stripped, co-registered, ~1 mm isotropic input like BraTS; RC and NETC are the least accurate classes; no clinical validation; volume differences between visits can be within the segmentation error. Do not use for diagnosis, treatment planning or follow-up of patients, and do not upload identifiable patient data to the demo.

License

The example scans and reference segmentations are derived from MU-Glioma-Post and are redistributed under CC BY 4.0 with attribution (see NOTICE.md). The license of the weights and code is set by the repository owner (license: other in the metadata above: edit it before making the repo public).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support