AIOmarRehan/medtrace-brats-segresnet

MEDTRACE brain tumour segmentation (SegResNet)

3D tumour compartment segmentation from four co-registered MRI sequences. Trained on BraTS 2023 GLI and published as part of the MEDTRACE longitudinal workstation, which deliberately does not measure with it. Why.

Research prototype. Not a medical device. Not for diagnosis, treatment planning, or any clinical decision. Not clinically validated. It has been measured against one annotation protocol on one dataset, which is agreement, not clinical accuracy.

Results on the held-out test split

186 cases from 169 subjects, split by subject so no subject appears in two splits.

Region	Dice mean	Dice median	HD95 median	Sensitivity	Precision
Tumour core	0.908	0.956	2.00	0.917	0.917
Whole tumour	0.922	0.949	2.45	0.925	0.924
Enhancing tumour	0.852	0.898	1.41	0.883	0.849

Empty regions follow the BraTS convention: with empty ground truth, Dice is 1 if the prediction is also empty and 0 otherwise, rather than being averaged away as NaN.

Use the test figures, not the validation ones. Epoch selection and post-processing thresholds were both tuned on the validation split, so validation numbers are optimistic by construction.

Files

File	Size	Contents
medtrace_seg_best.pt	72 MB	Selected weights, epoch 32
medtrace_seg.torchscript.pt	72 MB	TorchScript export, no MONAI needed to run
model_card.json	6 KB	Full machine-readable card: architecture, contracts, metrics, environment

Input contract

Channel order is not recoverable from the weights. The wrong order gives wrong output and no error.

channels:       [t1c, t1n, t2f, t2w]
normalisation:  per case, per channel, zero mean unit variance over NON-ZERO voxels only
spacing:        1.0 x 1.0 x 1.0 mm
preprocessing:  skull-stripped and co-registered, as BraTS 2023 distributes it
cropping:       non-zero bounding box of the summed channels, 4 voxel margin
patch size:     128 x 128 x 128
inference:      sliding window, overlap 0.5, gaussian blend

Output contract

Three overlapping channels with independent sigmoid activation, not a softmax over classes.

Channel	Meaning
TC	Tumour core: necrotic core + enhancing
WT	Whole tumour: necrotic core + oedema + enhancing
ET	Enhancing tumour

Thresholded at 0.5, with enhancing tumour suppressed below 200 voxels. To convert to BraTS integer labels, write WT as 2, then TC as 1, then ET as 3, in that order: the regions overlap, so the write order is what produces the correct nesting.

Architecture and training

monai.networks.nets.SegResNet, 18.8M parameters, 4 input channels, 3 output channels, init_filters=32, blocks_down=[1,2,2,4], blocks_up=[1,1,1], dropout_prob=0.2.

DiceFocalLoss with AdamW at lr 2e-4, cosine annealing, AMP float16, on a Tesla T4. 37 epochs completed of 60 planned; epoch 32 selected on mean validation Dice. Patch sampling was 80% centred on whole tumour, 20% uniform. Augmentation was random axis flips and intensity scale and shift within 10%. Seed 20260813.

MONAI 1.6.0, PyTorch 2.10.0+cu128.

Limitations

• Trained on pre-operative adult glioma only. Post-treatment appearances, including resection cavities and radiation change, are not represented.
• Requires all four sequences. Behaviour with one missing is untested.
• Assumes BraTS preprocessing: skull-stripped, co-registered, 1 mm isotropic.
• Agreement with one annotation protocol on one dataset is not a measure of clinical accuracy.

A note on how MEDTRACE uses this

MEDTRACE does not measure with this model, hosted or local. The hosted demo reads the expert-corrected segmentations that ship with RHUH-GBM; the local build reads DeepBraTumIA's masks on LUMIERE.

The reason is measured rather than cautious. Run against DeepBraTumIA on 12 randomly chosen LUMIERE studies, this model reaches 0.923 median Dice on whole tumour but 0.486 on enhancing tumour, below 0.5 in 6 of the 12 — and enhancing tumour is the single compartment MEDTRACE reports change on. Split by how much enhancement is present, the weakness is specific rather than uniform: 0.861 median Dice where enhancement is bulky (>= 5 cm3, n=5) against 0.193 where it is small (< 5 cm3, n=7), overestimating volume roughly fourfold in the small group. Post-operative brains are outside this model's training distribution, and on small lesions a few tenths of a cm3 is what decides whether progression is reported.

Publishing the model and measuring with it are separate decisions. Only the first is justified by the numbers above.

Training data and required citations

Trained on the BraTS 2023 GLI challenge training split, the only labelled split. The BraTS data use agreement grants use and requires citation; it does not grant redistribution, so no BraTS imaging appears in this repository or in any MEDTRACE artefact.

Menze, B. H., et al. (2015). The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Transactions on Medical Imaging, 34(10), 1993-2024. https://doi.org/10.1109/TMI.2014.2377694

Bakas, S., et al. (2017). Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific Data, 4, 170117. https://doi.org/10.1038/sdata.2017.117

Baid, U., et al. (2021). The RSNA-ASNR-MICCAI BraTS 2021 Benchmark on Brain Tumor Segmentation and Radiogenomic Classification. arXiv:2107.02314. https://doi.org/10.48550/arXiv.2107.02314

Related

• Space: AIOmarRehan/medtrace
• Dataset: AIOmarRehan/medtrace-rhuh-gbm-derived
• Code: github.com/AIOmarRehan/medtrace