BrowserQC

Automated MRI quality control, entirely in your browser — no upload, no server. Drag in a NIfTI image (or DICOM files or folders — with several series you pick one) and BrowserQC runs on its own: it segments the brain, shows the parcellation over your scan, and reports MRIQC-style image-quality metrics.

Live demo: browserqc.org.

How it works

Everything runs in WebAssembly + WebGPU (WebGL2 fallback) on your machine, so your images are never shared with the cloud. When an image loads (on startup and on every drag-and-drop):

1. Segment — @brainchop/mindgrab runs a brainchop model picked from the Model menu: mindmap PVE (default), GM/WM/CSF partial-volume fractions, the closest match to MRIQC; 16chan18cls (fast) or mindmap (24-channel), each parcellating the brain into 17 gray/white-matter and subcortical regions; or mindsnap (24-channel), 103 Desikan-Killiany cortical and subcortical regions. The mindgrab brain mask also runs, so QC counts extra-cerebral CSF as MRIQC does. Conform (256³ 1 mm), inference and back-projection to the native grid all happen inside the WebAssembly module — on WebGPU where available, WebGL2 otherwise (so no WebGPU is required).
2. Overlay — the labels (or fractions) arrive on the native input grid and are drawn as a colour overlay on the original scan (adjust with the Opacity slider). View → Background swaps to MRIQC's noise view (inverted viridis, windowed so artifacts in the air stand out); Right drag adjusts contrast or pans (wheel zooms). 2D slices show raw voxels (nearest-neighbour).
3. Quality control — niimath computes MRIQC-style anatomical image-quality metrics from the scan and its segmentation (with PVE, each voxel weighted by its tissue fraction, as MRIQC does), shown in the side panel:
   • CJV — coefficient of joint variation (noise + intensity non-uniformity); lower is better
   • CNR — contrast-to-noise (GM vs WM, over tissue + air noise); higher is better
   • SNRd — Dietrich SNR, normalised by background (air) noise; higher is better
   • FBER — foreground-background energy ratio; higher is better
   • SNR — signal-to-noise, averaged over tissues; higher is better
   • WM2MAX — white-matter median ÷ P99.95 intensity
   • EFC — entropy focus criterion (ghosting / blur); lower is better
   • Tissue composition — CSF / GM / WM as % of intracranial volume and absolute volume

Rendering uses NiiVue; DICOM import uses dcm2niix. A DICOM series' own sidecar becomes the report's bids_meta.

Metric names and definitions follow MRIQC, so outputs can be diffed against it directly. A Save button writes the metrics as MRIQC-style JSON (the CLI writes the same file); drop a BIDS sidecar .json alongside the image and it rides along as bids_meta. Rate opens MRIQC's manual rating widget (quality slider Exclude → Excellent, artifact checklist, comments, rater confidence); its Save writes <image>_rating.json with MRIQC's keys (dataset, subject, rating, artifacts, time_sec, confidence, comments), unlocking on your first edit at least 10 s after the image is displayed, as MRIQC does. Share publicly is disabled: there is no rating server.

This is a fast approximation of MRIQC, not a reimplementation: it uses a deep-learning parcellation (or its CAT-lite fractions) rather than MRIQC's Atropos partial-volume maps, but intensities follow MRIQC's pipeline: niimath clips them to 0–255 as niworkflows does, removes the bias field with N4 (a port of ANTs N4, weighted by the brain mask), and scales them so the eroded-WM median is 1000. Expect the same ballpark and the same ranking, not the same numbers. Not a substitute for MRIQC's normative values.

Develop

Requires Node ≥ 22 (npm run test:unit imports the TypeScript sources directly via Node's type stripping).

npm install
npm run dev         # vite dev server (http://localhost:8091)
npm run build       # typecheck + production build to dist/
npm run preview     # serve the production build
npm run test:unit   # node --test: BIDS-sidecar state machine, CLI wiring
npm run test:e2e    # build, then a headless-Chrome smoke of the auto-run path (needs Google Chrome)
npm run test:dicom  # build, then drop a many-series DICOM folder (needs the reproin sample)
npm run test:parity # build, then compare the web app with the native CLI (needs the native tools)

Command line

The same pipeline without a browser or Node — Python 3.8+ standard library only — on the native tools: brainchop-mindgrab and brainchop-<model> (brainchopC releases) and niimath newer than v1.0.20260926 (with --qc --pve --mask and N4; not yet released) on PATH (or in $BROWSERQC_BIN):

python3 cli/qc.py --in T1.nii.gz --out qc.json [--model mindmap-pve|16chan18cls|mindmap|mindsnap] [--bids sidecar.json]

It writes the report the page's Save does. npm run test:parity checks the two agree on the bundled image: segmentation and brain-mask Dice ≥ 0.999, median metric difference ≤ 0.02 %, worst ~0.4 % (BROWSERQC_T1=… npm run test:parity checks another image).

License

BSD-2-Clause. niimath comes from the @niivue/niimath npm package (BSD-2). Segmentation comes from the @brainchop/mindgrab npm package (MIT), a C11 MeshNet reimplementation compiled to WebAssembly + WGSL.

Links

This live demo already provides several of the core measures used by MRIQC (with more to come).

• MRIQC documentation
• Esteban et al. (2017) MRIQC: Advancing the automatic prediction of image quality in MRI from unseen sites. PLoS One PMID: 28945803