SSLMRI · MSSEG-2 · MICCAI 2024
AI-assisted MS lesion detection, visualization, and tracking.
NeuraSpec helps neurologists detect, volumetrize, and track Multiple Sclerosis lesions in brain and spine MRI — with an interactive 3D viewer and longitudinal progression reports.
Brain and spine MRINIfTI and DICOMDoctor, patient and admin roles
Three jobs, in one pass.
A neurologist reading MS follow-ups spends most of the appointment counting lesions by eye. NeuraSpec does the counting and the measuring, and hands back a picture worth arguing with.
Automatic detection
Deep-learning lesion segmentation trained on the MSSEG-2 dataset, based on the SSLMRI architecture.
ShippedInteractive 4-pane viewer
Sagittal, coronal, axial, and 3D volumetric reconstruction — powered by Cornerstone3D.
ShippedProgression tracking
Compare studies side-by-side, see new vs. resolved lesions, and generate PDF reports.
Phase 2 — in designScan in, report out.
Five steps from a folder of DICOM slices to something a patient can read. The first three are live today; the last two are the next phase.
Upload
A doctor drops a DICOM series into the app, against a patient record they own.
Analyze
A PyTorch model runs on a GPU server and returns a per-voxel lesion segmentation.
Visualize
Cornerstone3D renders four synchronized panes with the lesion mask laid over the scan.
Compare
Two studies align into an overlay diff — new lesions red, resolved green, persistent yellow.
Report
A PDF summarizing volumetric trend, lesion counts, and the doctor's own notes.
See NeuraSpec in action.
Walkthrough of an MS study from upload to AI-generated progression report.
How it's built
The model is the borrowed part and the plumbing is the built part — inference runs on hardware I own, because a free tier will not hold a segmentation network.
- Segmentation follows the SSLMRI architecture (Tahghighi et al., MICCAI 2024), trained on MSSEG-2.
- Inference is self-hosted on a Linux GPU box, so scan volumes never go to a third-party API.
- Three roles, three surfaces: doctors upload and correct, patients read, admins configure.
- Frontend
- React, Vite, TypeScript, Tailwind + shadcn/ui
- Imaging
- Cornerstone3D — 4-pane MPR and volume rendering, NIfTI and DICOM
- Backend
- FastAPI (Python 3.11), SQLModel, Alembic migrations
- ML
- PyTorch + MONAI, SSLMRI architecture, MSSEG-2 dataset
- Database
- PostgreSQL on Supabase
- Hosting
- Netlify for the client; self-hosted Linux GPU server behind nginx and Let's Encrypt for inference
- Status
- Phase 1 complete — auth, upload and the 4-pane viewer. Comparison and PDF reports are Phase 2.