MRI Recurrence
Risk ModelBREAST IMAGING GROUPResearch only · Hosted research inference

POST-TREATMENT MULTIMODAL MODEL

Multimodal insights.
Recurrence risk estimates.

Combine DCE-MRI, clinical variables and reports with the full post-treatment model.
Uploads are sent to this Hugging Face hosted service for inference. Use only authorized, de-identified research data. NIfTI inference files are removed after the request. DICOM preparation sessions expire after 20 minutes of inactivity; hosting-provider policies also apply.

MRI, report and clinical modalitiesThree modalities. One integrated model.
Model statusChecking…No access key required · Public research trial
Explore with a sample

Load a synthetic software-testing sample. Real patient examples are disabled.

01 Upload MRI and tumour mask

Upload NIfTI files registered, resampled and cropped using the original study workflow. All three files must share dimensions and spatial alignment. For raw DICOM ZIP, use the preparation workflow above.

Total upload limit: 100 MB. Temporary images are deleted after inference.

3D imaging preview

Reconstruct the uploaded tumour mask and inspect DCE 2 slices. Server-side processing; preview does not run or modify the prediction model.

Select MRI files and a mask, then build a preview.

02 Clinical information

Select or enter values directly; keep unavailable fields as “Unknown”. Do not enter patient names, IDs or other identifying information.

Pre-treatment stage

Post-treatment stage

Original model encoding is retained: post-treatment T4 and N2/N3 are treated as missing, not as N0.

Receptor data

Enter the original study E/P/H values (0–100); leave blank if unknown. The original model uses 0, 1–99 and 100 categories, not standard clinical positivity thresholds. Do not enter a HER2 IHC score directly as the H value.

Pathology and family history

Treatment information

“Unknown” and “No” treatment entries are both encoded as 0 in the original model. Provide known information wherever possible.

03 Report text

RadioBERT accepts up to 512 tokens. Reports in unvalidated languages may affect results.