Browser-based cardiac MRI segmentation — all processing runs locally
No data leaves your device MobileUNetV2 (2.09M params, 8 MB) — trained with knowledge distillation
How it works: The cine segmentation model (MobileUNetV2, 8 MB) auto-loads on page visit.
Upload DICOM files, and the app parses headers, normalizes images, resamples to 224×224,
runs inference via ONNX Runtime Web (WebAssembly), and displays segmentation overlays.
All computation happens in your browser — no data is transmitted to any server.
Supported formats: Cine (multi-phase 2D) or Scar/LGE (single-phase 2D).
Select all DICOM files belonging to one series.
Output: Click "Download Results" to get a ZIP with NIfTI volumes, CSV measurements, and a README.
Try it: Click "Cine Sample" above to download a sample cine DICOM series
(1 short-axis slice, 25 phases). Unzip, then select all .dcm files in the DICOM file picker.
Model: MobileUNetV2 with knowledge distillation (T=3.0, W=0.99). Trained on ACDC + M&Ms datasets
(633 subjects, 4 scanner vendors). Mean Dice: 0.883 (LV cavity: 0.921, LV myocardium: 0.851, RV cavity: 0.875).
Reference: Feng X, Huang G, Meyer CH. CPU-based fast cardiac cine MRI segmentation
using lightweight CNNs and knowledge distillation. NMR in Biomedicine (under review).