brain-mets-autocontour-t1post

Brain Metastasis Auto-Contouring (T1 post-contrast MRI)

Model card only — weights are not distributed here. See Weight availability.

A 3D nnU-Net trained to segment contrast-enhancing brain metastases on post-contrast T1 MRI, built as part of a working radiotherapy auto-contouring tool that exports DICOM RT Structure Sets directly importable into a treatment planning system.

Developed by a physician with three years of radiation oncology training, working as an ML engineer.

Intended use

Research use only. This model is not a medical device, has not undergone independent clinical validation, and must not be used to inform patient care.

Intended for research on automated target delineation for brain metastasis radiotherapy planning. Any clinical use would require independent validation, regulatory clearance (MDR in the EU / FDA in the US), and a qualified clinician reviewing every contour.

Model details

Performance

Read the Dice figure correctly. It is nnU-Net's own held-out-fold metric computed during training, not independent validation against expert contours on unseen patients. No such independent evaluation has been performed, and no clinical Dice should be attributed to this model until it has.

Latency excludes DICOM→NIfTI conversion, skull stripping and RTSS export; end-to-end wall clock was not instrumented.

Training data

UCSF-BMSR — University of California San Francisco Brain Metastases Stereotactic Radiosurgery MRI Dataset, release 1.3. 560 brain MRIs from 412 patients with expert annotations of 5,136 metastases.

Rudie JD, et al. The University of California San Francisco Brain Metastases Stereotactic Radiosurgery (UCSF-BMSR) MRI Dataset. Radiology: Artificial Intelligence (2024). 10.1148/ryai.230126

Weight availability

Weights are deliberately not published here.

The UCSF-BMSR data use terms are custom and specify non-commercial, attribution, no redistribution, no re-identification. Whether publishing weights derived from that data constitutes redistribution is not settled by the licence text available to us, so the conservative reading applies: nothing derived from the dataset is redistributed until the data owners confirm otherwise.

This card documents the model so the work is inspectable and reproducible in description. It is not a workaround — no weights, no images, no data.

Researchers who want to evaluate the model can contact the author; a request for explicit permission from the dataset owners is the intended path to releasing weights.

Limitations

• Segments GTV only — no CTV or PTV margins.
• Trained on a single public source; no external-cohort validation.
• The 5-fold ensemble is internal validation, not independent evidence.
• Performance on scanners, sequences or populations unlike the training cohort is unknown and should be assumed worse.
• Small lesions are the known hard case for this task and were not separately characterised here.

Related work

Part of a portfolio of medical-imaging ML projects covering head-and-neck GTV segmentation benchmarking, breast BI-RADS assessment with external validation, and benchmark-contamination auditing.

Author

Ricardo Oyarzún-Silva — physician (three years of radiation oncology training) and ML engineer. ORCID 0009-0000-3086-9047