CALMaR

Co-designed Automated Lesion Mapping and Reporting for stroke lesions, running entirely in the browser:

1. Auto-segment a stroke lesion from a structural MRI (ONNX model in a Web Worker).
2. Normalize the patient brain to MNI152 with a deep-learning registration model.
3. Choose an Atlas and compute lesion overlap with either Yeo 7 networks or Schaefer 400 parcels.
4. Combine precomputed normative functional-connectivity maps into a lesion-network map. Yeo uses the supported 7-channel development_fmri pack; Schaefer uses the supported public N=155 development_fmri pack through the lazy-shard loader.
5. Threshold and visualize on top of MNI.

No backend. No data upload. Hosted on GitHub Pages.

The app opens on a short start page that explains the local-first workflow: load a structural T1, review the lesion mask, then map and export results. Patient images, masks, voxel values, screenshots, and generated outputs stay on the user's computer; public atlas and model assets may be fetched by the browser when needed.

Status

Selectable atlas implementation (current) — the run controls now include a visible select field labelled exactly Atlas.

• Schaefer 400 parcels is the default atlas. Yeo 7 networks remains fully supported and selectable for compatibility and comparison.
• Schaefer 400 parcels is available for direct lesion overlap via the official Schaefer2018 400-parcel, 7-network, 2 mm atlas. Result tables and CSV export use parcel labels instead of Yeo network names; Schaefer display labels omit the leading 7Networks_ prefix.
• The atlas registry in web/js/app/atlas-options.js maps each option to its overlap atlas, connectome asset, FC weighting mode (network or parcel), colormap, affected-map label atlas, and optional functional-profile asset.
• Schaefer FC generation is implemented as scripts/build_schaefer400_connectome.py: it fetches public Nilearn development_fmri subjects, computes 400 parcel-seed group t-stat maps, emits float16 row-major shards, writes a 4 mm Schaefer label companion, and can upload to the HF dataset when HF_TOKEN is available. The manifest points at the uploaded 10-shard, 118 MB pack and counts only the small index as cold-load because the browser fetches only shards containing lesion-hit parcels.
• Functional profile panels work for both atlas choices. Yeo uses the compact Yeo7 Neurosynth/NiMARE profile asset directly; Schaefer uses a parcel-wise NiMARE ROI decode over all 400 Schaefer parcels. Neither Yeo nor Schaefer is described as a dedicated language-area atlas.
• Spatial guardrails tag app-created NIfTI files as native-t1, mni160, or atlas:<assetId> and assert those spaces before computation and viewer overlay steps. This prevents atlas-grid masks, native review masks, and MNI160 registration products from being mixed silently.

Phase 1 complete (v0.1.0) — manual-mask Yeo 7-network overlap. Drop a binary lesion mask aligned to MNI152NLin2009cAsym 2mm, click "Compute overlap", get a per-network table with voxel counts, % of lesion, an inline magnitude bar, and a CSV export. Voxels falling outside the Yeo brain mask are surfaced as a warning. The Yeo7 atlas is fetched live from sbollmann/lnm-webapp-models on Hugging Face and cached client-side.

Phase 2 complete (v0.2.0) — auto brain extraction + lesion segmentation on T1.

Drop a structural T1; the app:

1. Runs SynthStrip brain extraction automatically in a module worker (~7–10 s on M-series, WASM EP, single-pass). Result is rendered as a translucent green overlay and downloadable as lnm-brainmask.nii. Model: manifest asset lnm-synthstrip, an ONNX FP32 export of FreeSurfer SynthStrip (SynthStrip-Hoopes2022-Apache-2.0) ported from neurodesk/vesselboost-webapp. See Brain extraction model provenance for the exact hosted asset, checksum, and browser wrapper contract.
2. On click of "Run lesion segmentation", runs SynthStroke baseline (the closest available openly-licensed model for ATLAS-2-style chronic stroke on T1; 3D MONAI UNet, MELBA 2025, MIT). ~5 s per pass on M-series. Result is rendered as a translucent red overlay and downloadable as lnm-lesion.nii. Sliding-window 128³ patches, threshold 0.4, min cluster 30, overlap 0.25, no TTA. See Lesion segmentation model provenance for the exact upstream model, conversion script, and browser runtime differences from the upstream 192³/TTA recipe.

Phase 4 complete (v0.4.0) — Yeo7 group functional-connectivity weighted sum. After running "Compute overlap" on a manual MNI 2 mm lesion mask, click "Compute network map" — the orchestrator fetches a 30 MB Yeo7 FC pack (7 brain-wide t-maps, computed from 30 ADHD-200 subjects via scripts/build_yeo7_connectome.py), weights each network's t-map by the lesion's share of that network, and emits a Float32 NIfTI on the Yeo7 atlas grid (99×117×95 2 mm). Output renders as a red-yellow overlay and downloads as lnm-network-map.nii. Pure main-thread JS — no worker round-trip; the math is a per-voxel linear combination via web/js/modules/fc-weighted-sum.js.

Phase 3 complete (v0.3.0) — deformable MNI registration via SynthMorph (Hoffmann 2022, Apache-2.0). Click "Run MNI registration" on a 160×160×192 1mm structural T1; the app fetches the SynthMorph ONNX (81 MB) + the lnm-mni160 reference (8 MB), downsamples the source/reference pair to the browser graph's 48×64×80 grid, runs the SVF-only sub-network in the worker, then performs scaling-and-squaring SVF integration + 24×32×40→160×160×192 upsample + the spatial warp in pure JS (web/js/modules/registration.js). WebGPU execution provider when available; falls back to WASM.

The lnm-yeo-auto pipeline declares the full T1 → SynthStrip → seg → register → MNI Yeo overlap chain. Stages run individually for now (each is a button click); a one-shot "auto" runner that also resamples the warped lesion onto the MNI152 2mm Yeo grid is a follow-up polish slice.

Experimental notes: SynthMorph's deformable head expects roughly- MNI-aligned input. Without an upstream affine pre-step (FSL FLIRT, ANTs antsRegistrationSyNQuick), deformable registration on raw clinical T1 may not converge well. Inputs must be exactly 160×160×192 at 1mm; the orchestrator surfaces a clear error otherwise.

Phase 39 (v0.17.0) — assume raw T1; drop visible manual-mask input.

The Input section used to show two file inputs: "Structural T1" and "Or: pre-computed lesion mask (Yeo 2 mm grid)". Clinical use is T1-only — the second input was researcher-mode noise. Simplified:

• The lesion file input moves under Advanced as "Researcher mode: pre-computed lesion mask" with a clear note that the mask must already be on the selected atlas grid.
• The Input section now shows ONE input (T1) plus a help line explaining the auto chain handles everything from raw T1.
• lnm-yeo-only and lnm-network-map pipelines are flagged hidden: true so they drop out of the dropdown. Visible pipelines: lnm-yeo-auto (the default) and lnm-segment-only. Hidden pipelines remain reachable via setLesion() auto-promote when a researcher loads a Yeo-grid mask through the Advanced input or when the browser smoke drives #lesionFileInput directly.
• Static <select> fallback option in index.html is now lnm-yeo-auto (the auto chain) instead of lnm-yeo-only.

Test contract pinned: test:tasks asserts both manual-mask pipelines are hidden + lnm-yeo-auto is visible. test:html asserts the static fallback is now the auto pipeline.

Phase 38 (v0.16.2) — N=155 development_fmri connectome + smoke verify.

• Connectome upgrade: ADHD-200 N=40 → development_fmri N=155 (~4× sample size). scripts/build_yeo7_connectome.py parameterised with a --dataset {adhd, development_fmri} flag; same processing pipeline (per-subject Yeo7 ROI mean → Pearson r-map → Fisher-z → group t-stat). New pack uploaded to HF as connectomes/yeo7_fc_pack_dev155.bin under cacheKey yeo7-fc-pack-development-n155-v1. Manifest entry repointed; old N=40 pack still accessible via its prior URL for users pinning the older cacheKey.

  Group t-stat range: [-13.47, +30.72] (vs N=40's [-13.85, +19.74] — peak +stat ~55% higher with the larger sample, as expected for a fixed-effect group analysis).

• Browser smoke verified at v0.16.1: 6/6 phases green (1c.4 + 8 + 2a.1.5 + 2a.2.5 + 3.7 + 10). Confirms Phases 34/35/36/37 didn't regress anything user-visible. SynthMorph EP=wasm (headless Chromium + swiftshader doesn't satisfy ORT's WebGPU device requirements; degraded path fires as expected).

Phase 35 (v0.16.1) — controller + UI module behavior tests.

The Phase 33 audit flagged 4 controllers + 3 UI modules as source-grep-only. This phase replaces that with real coverage:

• test:viewer-controller — fake NiiVue (records every call); pins the Phase 4 silent-regression bug: loadBaseVolume MUST call nv.loadVolumes([single]) (1 entry), loadOverlay MUST use nv.addVolumeFromUrl(...). A future "simplification" back to loadVolumes([base, ...overlays]) (which leaves overlay LUT uninitialised in 0.68.x) fails immediately. 9 cases.
• test:file-io-controller — fake DOM File events; pins NIfTI vs DICOM detection (case-insensitive .nii/.nii.gz), the mixed-list picker, dispatch via onFileLoaded, drop-item routing. 9 cases.
• test:inference-executor — fake Worker constructor; pins the message protocol (init/load/run-synthstrip/run-register/etc.), stageData routing into this.results, step-complete state transitions, error-path onError + state cleanup, cancel() termination + idle-no-op, removeResult/clearResults, volume-info + complete callbacks. 11 cases.
• test:ui-modules — fake DOM; pins ConsoleOutput log/clear/ missing-element no-op, ProgressManager setProgress paints width: NN% + reset-to-0, ModalManager open/close/toggle/isOpen + overlay-click-to-close (target===modal closes; click on a child doesn't). All three modules also tested for graceful behaviour when the bound DOM id doesn't exist.

Test surface: 24 → 28 distinct suites in npm test. Closes the 7 zero-coverage modules from the Phase 33 audit.

Phases 34 + 36 + 37 (v0.16.0) — auto-prealign + PCA orientation fix + download progress.

• Phase 34: auto-prealign in runFullPipeline — clinical T1 workflow now requires one click, not three. The lnm-yeo-auto pipeline gains a prealign stage between brain-extraction and inference-pipeline. prealignToMni160({ skipIfAligned: true }) is idempotent: probes the structural's dims and no-ops when already at 160×160×192. New prealign module added to IMPLEMENTED_MODULES and _runStage's case dispatch.

• Phase 36: PCA 180° fix via NIfTI affine prior — the Phase 33 audit flagged that PCA on a brain mask gives the principal axes up to sign, so a clinical T1 acquired upside-down would silently prealign to a mirror-image brain. Resolution: trust the source NIfTI affine. After PCA, project each column from source-voxel to source-world space, pick column signs that make R_world's diagonal positive (PCA axes align with world axes the source declares). When re-enforcing det = +1, flip the most ambiguously-oriented column (smallest |R_world[k][k]|) instead of always column 2 — which is what the previous implementation did and undid the sign correction. test:prealign-pca-orientation flipped from documented-limitation to hard assertion.

• Phase 37: download progress for FC pack fetch — the FC pack (~30 MB cold) used to download silently from Hugging Face during network-map runs. Now loadConnectomeFromManifest accepts an onProgress callback that the orchestrator wires into the existing status-bar progress. fetchCacheFirst opts into a streaming-tee path when a callback is supplied (cache hits skip the streaming). Exception in onProgress is swallowed — best-effort. Worker model fetches already had progress (Phase 19); this closes the atlas-side gap. Two new test cases added to test:atlas-loader-cache.

  Honest correction: the original Phase 37 plan claimed lazy-loading ORT WASM would drop first-page-load from 38 MB to 3 MB. On further inspection this was already the case — the worker (and therefore ORT) is created lazily; the 35 MB WASM only downloads on first inference. Pivoted Phase 37 to download-progress UX, which is the actual user-facing gap.

Phase 33 (v0.15.1) — test-suite audit + tightening.

Audit found three loose thresholds and four uncovered modules. Fixed:

Tightened thresholds:

• test:resample-parity Dice: ≥ 0.95 → exact = 1.0. The math is bit-exact for an aligned-grid 2× nearest-neighbour roundtrip.
• test:real-data-bridge per-network tolerance: flat ±25 voxels → max(15, 0.05 × expected). Stops Visual (n=125) from getting 20% slack while Limbic (n=1904) gets 1.3%.
• test:synthstrip-parity coverage: 10–95% → 18–60%; centroid drift: <15 voxels → <10 voxels. Catches "model output is half a hemisphere" while accommodating the MNI template's natural inferior offset.

4 new behavior tests (replace source-grep coverage):

• test:app-behavior — partially-instantiates LesionNetworkMappingApp with stubbed browser globals, replaces _runStage with a recording spy, and asserts: stages dispatch in declared order, stage exception halts the chain, precondition gates fire, _runStage rejects unknown modules, _autoPromotePipeline only fires before user manual pick. Catches a regression where the for-loop accidentally returns after stage 1.
• test:nifti-writer — round-trips Float32 + Uint8 phantoms through writeNifti1 and nifti-reader-js, asserts byte-equal voxel values, dim/spacing/affine round-trip, NIfTI-1 magic bytes. Was zero-coverage despite being load-bearing for every download button.
• test:atlas-loader-cache — exports fetchCacheFirst and pins the Phase 4 silent-fail bug fix: bare cacheKey strings parse as URL schemes and crash Cache.put; the fix folds them into a URL fragment. Tests warm-cache short-circuit, non-fatal cache failure, and HTTP-error propagation.
• test:prealign-pca-orientation — DOCUMENTS the 180° ambiguity limitation: PCA covariance is identical for upright and flipped acquisitions, so principalAxisAlign produces mirror-image outputs for the same anatomy. The test confirms equal mass / equal skew magnitudes on both poses; the orientation sign mismatch is logged but not hard-asserted (a 3rd-moment fix would interact poorly with the det = +1 enforcement; resolution needs an A/P/L/R anatomical prior or accepting MNI's left-handed FSL convention).

Test surface: 29 distinct test scripts, 21+ in npm test (was 21). 6 browser smokes + 4 opt-in heavy parity tests unchanged.

Phase 32 (v0.15.0) — UI cleanup for clinical use.

Audit found the sidebar had grown to 3 stacked sections × 6+ buttons each and the toolbar carried 9 controls of which 6 had no JS binding (carried over from the SCT scaffold). Redesigned:

• Sidebar collapses to three numbered panels:

  1. Input — file inputs only.
  2. Run analysis — single primary button. The 7 per-stage controls (brain extraction, prealign, lesion seg, registration, warp, Yeo overlap, network map) move into an <details> "Advanced" disclosure that's closed by default.
  3. Results — overlap table + threshold panel + 3 essential downloads (CSV, network map, thresholded mask). Brain/lesion mask downloads move into an "Intermediate downloads" disclosure.

  All 27 DOM IDs preserved (the per-stage buttons just live inside <details> now), so every existing JS binding + the test_index_html.mjs lockdown still work.

• Toolbar trimmed: dropped 6 dead controls (windowMin/Max, rangeMin/Max, rangeSelected, resetWindow, inputVisibilityToggle, downloadCurrentVolume, screenshotViewer) — none had JS bindings. Kept: view tabs, overlay opacity (now visible by default; was hidden), crosshair, smoothing, colorbar, base colormap.

• Section headers are now static labels (no longer click-to-collapse — the per-stage Advanced controls use native <details> instead). toggleSection script removed; .sidebar-section.collapsed CSS rules removed.

• Browser smoke: 4 tests broke (per-stage buttons inside closed disclosure → "not visible" to Playwright). Added openAllDisclosures(page) helper called after each page.goto; idempotent. Phase 8/10 (primary flow) untouched.

Phase 31 (v0.14.1) — smoke regressions fixed + docs refresh.

Ran the browser smoke for the first time since v0.10.0 and surfaced two issues I'd been carrying:

1. Pipeline auto-promote — Phase 15 made runFullPipeline iterate selectedPipeline.stages. Default selection lnm-yeo-only only had [parcel-overlap], so "Run full pipeline" with a manual mask only ran overlap. Phase 8 smoke timed out polling for thresholdedMaskFile.

   • setStructural() now auto-promotes the dropdown to lnm-yeo-auto (full auto chain), setLesion() to lnm-network-map (overlap + FC + threshold) when no structural is loaded.
   • Suppressed once the user manually picks a pipeline (_userPickedPipeline = true).
   • Result: Phase 8 smoke now passes in 4.3s (was 120s timeout).

2. Browser-runnable SynthMorph graph — the original 160×160×192 ONNX graph hit a multi-gigabyte first Conv3D activation in ORT WebGPU/WASM. The registered browser asset is now lnm-synthmorph-mni-48x64x80.onnx: same SynthMorph weights, smaller static graph, source/reference downsampled before ONNX, and displacement upsampled back to MNI160. npm run test:synthmorph-browser-model gates the activation budget, and Phase 3.7 smoke now requires browser registration completion.

AGENTS.md architecture entry for prealign.js added; test surface table refreshed (20 Node suites listed).

Phases 26–30 complete (v0.14.0) — autonomy push.

This batch closes work I'd previously framed as needing user input, after a re-review showed every blocker was softer than it sounded.

• Phase 26 — PCA principal-axis prealign web/js/modules/prealign.js now exports covarianceOfMask, jacobiEigen3x3, and principalAxisAlign. The prealigner rotates the brain's principal axes onto MNI canonical axes (forced right-handed) on top of the centroid match. Validated on synthetic rotated phantoms AND on the real ds004884 T1 — PCA lands the brain mask centroid at MNI160 voxel (79, 81, 96) within 1.5 voxels.

• Phase 27 — N=40 ADHD-200 connectome rebuild + HF upload Re-ran scripts/build_yeo7_connectome.py at the maximum N (40) supported by nilearn.datasets.fetch_adhd, uploaded as connectomes/yeo7_fc_pack_n40.bin (cacheKey yeo7-fc-pack-adhd200-n40-v1), bumped the manifest. New group t-stat range [-13.85, 19.74]; old N=30 pack remains accessible via its prior URL.

• Phase 28 — SynthMorph EP introspection Worker now tries WebGPU explicitly first, catches the failure, then retries with WASM, and logs SynthMorph EP=<name>. Smoke surfaces the chosen EP in test output. Catches a regression where a future ORT upgrade silently always picks WASM (which OOMs on the 4 GB heap).

• Phase 29 — PCA on real anatomical data New scripts/test_real_data_pca.mjs runs the Phase 26 pipeline on the ds004884 T1 with a quick intensity-threshold brain mask. Asserts positive eigenvalues, dominant principal axis, det(R)=+1, and the resampled centroid lands within 1.5 voxels of MNI center.

• Phase 30 — real-data Dice parity gate tests/fixtures/ds004884-mini/expected_yeo_overlap.json pins the network voxel counts for ds004884; the bridge test asserts max-abs diff stays under 25 voxels per network. Catches a silent shift in the resample / bridge / overlap chain that doesn't crash but moves the answer.

Phase 25 complete (v0.13.2) — manifest checksum verifier.

scripts/test_manifest_checksums.cjs walks every supported asset in web/models/manifest.json, locates its bytes in either web/models/_dev_cache/ (developer machine) or tests/fixtures/yeo7-mini/atlas.nii.gz (committed copy), and asserts the actual sha256 matches the manifest's declared one. Catches:

1. The committed Yeo7 fixture drifting from the runtime asset.
2. A model rebuild landing in _dev_cache without a matching manifest bump (cacheKey collision → browser serves stale cached bytes).
3. A manifest typo where filename / sourceUrl doesn't match the actual asset id.

Skips assets without a local copy (most CI runners don't fetch the ~200 MB of weights), but requires ≥ 1 verified asset so the gate doesn't silently no-op. Currently verifies 7 / 7 supported assets (including independent verification of the committed Yeo7 fixture) on a developer machine.

Phase 24 complete (v0.13.1) — real-data Yeo overlap in npm test.

tests/fixtures/yeo7-mini/atlas.nii.gz (73 KB, identical bytes to the runtime asset) committed alongside the existing ds004884-mini lesion fixture. scripts/test_real_data_bridge.mjs extended to load the atlas, run computeParcelOverlap + summarizeNetworkOverlap on the prealigned + Yeo-resampled lesion, and assert:

• ≥ 2 Yeo networks hit (real chronic strokes span vascular territories that cross network boundaries)
• network voxel sums + out-of-atlas voxels = total Yeo lesion count

Output for the committed ds004884 case (left-hemisphere chronic stroke): 7 networks hit; Limbic 11.6%, Default 10.5%, Frontoparietal 7.1% dominant; ~63% subcortical / outside the cortical Yeo mask.

Phase 21+22 complete (v0.13.0) — UX + deploy budget.

• Reset state button: New "Clear results" button in the Results section drops every intermediate file (overlap, network map, thresholded mask, brain mask, lesion mask, MNI lesion, network map raw data) + re-disables every download button + clears the table body and threshold summary. Structural file is retained so users can quickly re-run on the same input.
• Deploy-size budget: New scripts/test_deploy_budget.cjs walks the static deploy artifact (web/ minus _dev_cache) and tallies every supported manifest entry's sizeBytes for runtime fetches. Asserts static < 60 MB (currently 38) and total cold-load < 300 MB (currently 247). Locks the original plan's "<200 MB cold load" acceptance criterion against future model swaps.

Phase 18 complete (v0.12.1) — real-data bridge integration test.

The unit suites covered the resample math on synthetic phantoms and the prealign math on synthetic centroids; this slice connects them on real anatomical shape so a regression that fires only on a real lesion surfaces in CI. New scripts/test_real_data_bridge.mjs runs in npm test:

1. Decodes the committed ds004884-mini lesion mask (160×256×256 1mm real chronic stroke, ~131k voxels).
2. Runs centroidOfMask → applyAffineToVoxel → computePrealignAffine.
3. Resamples the lesion onto MNI160 1mm via resampleAffine(...).
4. Resamples again onto the canonical Yeo7 99×117×95 2mm grid.
5. Asserts: source 131k → MNI160 ~131k (within ±15% for a 1mm→1mm shift+flip), MNI160 → Yeo7 within ±50% of the expected ⅛ downsample, and the MNI160 centroid lands at voxel (80, 80, 96) within 1 voxel.

Phase 16 complete (v0.12.0) — in-browser affine pre-registration.

The SynthMorph deformable head requires its input at exactly 160×160×192 1mm AND roughly MNI-aligned. Real clinical T1s come in arbitrary dims, voxel sizes, and ACPC orientations, which previously meant external prep (FSL FLIRT / ANTs) before the auto chain could run. This phase ships a centroid-match prealigner in pure JS:

• New module web/js/modules/prealign.js exports centroidOfMask, applyAffineToVoxel, and computePrealignAffine. The destination affine places the source brain centroid at MNI160 voxel (80, 80, 96) under canonical FSL orientation (-x, +y, +z); when fed to resampleAffine(..., 'trilinear') it produces a 160³ 1mm prealigned T1 + brainmask in one pass.
• New orchestrator method prealignToMni160() and a new "Pre-align to MNI 160³ 1mm" button in the Lesion section. Runs SynthStrip first if the brainmask is absent, then resamples T1 + brainmask in pure JS, clears stale downstream state, and rerenders the viewer.
• Test suite: scripts/test_prealign.cjs covers the math (cube centroid, identity / scaled+translated affines, round-trip from MNI voxel back to source world centroid). Wired into npm test.

Limitations: centroid match only — no rotation correction, no intensity-based optimisation. Works well for ACPC-aligned scans (modern T1s); rotated clinical acquisitions or scans with severe pathology may need a follow-up rigid pass (Phase 16 v2).

Phase 20 complete (v0.11.1) — CI/CD lock-down. Both .github/workflows/deploy-pages.yml and .github/workflows/release.yml were carried over from the SCT scaffold and broken on this fork (referenced Git LFS we don't use, web/models/*.onnx files that don't exist, and missing helper scripts get_version.sh / summarize_test_failures.cjs). Rewritten:

• LFS removed throughout (we fetch ONNX from Hugging Face at runtime).
• Deploy verification now checks web/wasm/ort-wasm*.wasm (the only weight-class artifact we actually ship).
• New scripts/get_version.sh parses VERSION out of web/js/app/config.js for the release-tag bump step.
• Release workflow simplified — runs npm test, uploads the log on failure, computes the next patch version, tags + creates the GitHub release, then the deploy workflow's workflow_run trigger redeploys production from the new tag.

Phase 19 complete (v0.11.0) — per-stage perf instrumentation.

runFullPipeline() now wraps each stage in performance.now() markers, logs a [perf] <stage> (<module>): NN.NN s line per stage, and emits a final === Pipeline complete in X === summary with the total runtime and stage count. Stage timings are also collected into window.app._perfStats for inspection (e.g. JSON.stringify(app._perfStats) in the dev console). No external deps; same suite stays green.

Phase 15 complete (v0.10.0) — pipeline-driven runFullPipeline.

• runFullPipeline() now iterates selectedPipeline.stages instead of hard-coding the chain. Selecting a different pipeline truncates or extends what runs:
  • lnm-yeo-only: overlap (manual mask).
  • lnm-segment-only: brain extraction → lesion seg.
  • lnm-network-map: overlap → FC → threshold (manual mask).
  • lnm-yeo-auto: brain extraction → seg → register → overlap → FC → threshold.
• New _runStage(stage) dispatches on stage.module. Unknown modules throw so a manifest typo surfaces immediately.
• Threshold module added to IMPLEMENTED_MODULES. Stage-level defaults are pushed into the threshold UI controls before applyNetworkThreshold() runs, so each pipeline can ship its own threshold preset.
• Test contract pinned: _runStage must have a case for every implemented module; runFullPipeline must iterate pipeline.stages.

Phase 14 complete (v0.9.1) — cancel-button wiring. The #cancelButton next to the status line now actually terminates the worker via executor.cancel(). Disabled state is driven from handleWorkerProgress (enabled while progress fraction is in [0, 1)) and handleStepComplete (disabled at end-of-step), matching the SCT-era behaviour for which the button was originally drawn.

Phase 13 complete (v0.9.0) — UX surface.

• Pipeline dropdown now lists every runnable pipeline declared in lnm-tasks.js, not just lnm-yeo-only. The Schaefer400 legacy pipeline (lnm-default) remains hidden: true; the user-facing Schaefer path is the Atlas selector.
• New isPipelineRunnable(pipeline) helper + Node test enforces the dropdown filter contract.
• About modal now shows the actual Config.VERSION instead of an empty placeholder.

Phase 8–11 complete (v0.8.0) — fixtures + smoke + docs.

• New 160×160×192 1mm fixture (tests/fixtures/lnm-auto-mini): MNI152NLin2009cAsym 1mm template + planted hypointensity sphere as a smoke-test stand-in for a real stroke T1.
• Browser smoke now covers both branches of the full pipeline button. Phase 8 (manual mask, ~15 s) and Phase 10 (auto chain, ~5 min cold) exercise the orchestrator end-to-end.
• AGENTS.md "Test surface" + "Key Conventions" sections rewritten from the SCT scaffold to the current LNM invariants (module worker lazy nifti load, Cache Storage URL fragment trick, NiiVue addVolumeFromUrl call shape, SynthMorph 160³ hard check, bridge/runFullPipeline contract, threshold UI live-update).

Phase 7 complete (v0.7.0) — polish + parity guard.

• Citations modal updated with Yeo, SynthStrip (Hoopes 2022), SynthStroke (Chalcroft 2025), ATLAS v2.0 (Liew 2022), SynthMorph (Hoffmann 2022), ADHD-200, and the canonical lesion-network-mapping paper (Boes 2015).
• runFullPipeline() now detects a manually-loaded Yeo-grid lesion mask and skips segmentation/registration/bridge — going straight to overlap → FC → threshold so the manual flow is one click too.
• New test:resample-parity suite asserts a Yeo→MNI160→Yeo nearest-neighbor roundtrip preserves a 6³ phantom at Dice = 1.0 with no centroid drift, locking down the bridge math against future changes.

Phase 6 complete (v0.6.0) — end-to-end auto-pipeline. The native lesion mask produced by SynthStroke is now bridged onto the Yeo7 MNI 2 mm grid in two steps the orchestrator chains internally:

1. The worker applies the SynthMorph integrated displacement field to the F-order lesion voxels (stepWarpMask → 160×160×192 1 mm).
2. Pure-JS web/js/modules/resample.js performs an affine-aware resample (NIfTI sform → 4×4 inverse → per-voxel destination lookup) onto the Yeo atlas grid (99×117×95 2 mm), nearest mode for binary masks.

The result is adopted as this.lesionFile so a downstream runYeoOverlap → runFcNetworkMap → applyNetworkThreshold chain runs without any extra plumbing. A new "Apply registration to lesion" button exposes step (1)+(2); a "Run full pipeline" button chains brain extraction, lesion segmentation, registration, the bridge, Yeo overlap, FC network map, and threshold (defaults) in one click.

Phase 5 complete (v0.5.0) — thresholding UI + cluster cleanup. The "Network map" subsection now exposes a Threshold panel:

• Mode: absolute (slider is t-stat) or top percent of |voxels| (slider is 0–10 in 0.1% steps; 5 keeps roughly the strongest 5% and 0 keeps none).
• Symmetric toggle: |x| > T instead of x > T for positive/negative one-sided.
• Min cluster (voxels): post-threshold 26-connected component cleanup via the existing removeSmallComponents helper. The summary reports how many voxels were removed by this cleanup, so a value below the current connected component sizes is explicit rather than silent.
• A live summary line reports the survivor count; a Download thresholded mask button emits a Uint8 NIfTI binary mask (lnm-network-map-thresh.nii).

Pure JS in web/js/modules/threshold.js; the slider re-fires applyNetworkThreshold on every input change so the mask + summary stay in sync with the controls.

Attribution

The browser scaffolding (NiiVue viewer integration, ONNX Runtime Web worker pipeline, NIfTI/DICOM I/O, GitHub Pages deploy workflow) is adapted from neurodesk/spinalcordtoolbox-webapp. See THIRD_PARTY_NOTICES.md (added in Phase 1) for full credit.

Pipeline-specific dependencies (added incrementally):

• Brain extraction: SynthStrip (Hoopes 2022, Apache-2.0); manifest asset lnm-synthstrip, ONNX export ported from neurodesk/vesselboost-webapp.
• Lesion segmentation: SynthStroke baseline (Chalcroft 2025 MELBA, MIT); 3D MONAI UNet, T1.
• Registration: SynthMorph (Hoffmann 2022, Apache-2.0); UNet-only ONNX cut (layers 0–33); JS-side SVF integration + warp.
• Atlases: Yeo 2011 7-network cortical atlas; Schaefer 2018 400 × 7 networks (CC-BY).
• Connectomes: Yeo7 development_fmri N=155 group functional connectivity (computed by scripts/build_yeo7_connectome.py, Yeo7 ROI seed-to-voxel t-maps). Schaefer400 development_fmri N=155 group functional connectivity is built by scripts/build_schaefer400_connectome.py and served as 10 lazy float16 shards with a 4 mm Schaefer companion atlas.
• Functional profiles: Exploratory term associations from compact Neurosynth v7 / NiMARE profile assets. Yeo7 profiles are rebuilt offline with scripts/build_yeo7_function_profiles.py; Schaefer400 profiles are generated by scripts/build_schaefer400_function_profiles.py with a parcel-wise NiMARE ROIAssociationDecoder over all 400 Schaefer ROIs. Both are weighted in-browser by direct lesion overlap or thresholded connectivity-map effects.

Lesion segmentation model provenance

CALMaR uses exactly one default SynthStroke lesion-segmentation asset:

• Manifest id: lnm-stroke-lesion
• Runtime filename: models/lnm-stroke-lesion.onnx
• Source version: SynthStroke-baseline-MELBA2025-MIT
• Browser format: onnx-fp32
• Cache key: lnm-stroke-lesion-synthstroke-baseline-v1
• Size: 74518462 bytes
• SHA-256: 49f3d047a2299501791c7176cfda770d45d2f411a18ef41dd1105be4d0582cb8
• Hosted source: sbollmann/lnm-webapp-models/models/lnm-stroke-lesion.onnx

The ONNX asset was converted from the Hugging Face liamchalcroft/synthstroke-baseline model:

• Upstream model: SynthStroke baseline, not SynthStroke SynthPlus
• Upstream revision observed for provenance: b693a650026359705688fbce409219c4dbb5d6be
• Upstream files: config.json and model.safetensors
• Upstream config SHA-256: 2d9e7eb2ab4cb0a696ce6a845ad3123b77b31867f7ba74a271b81132cca38b1e
• Upstream weights size: 74468100 bytes
• Upstream weights ETag: d56c089e8c4bcc0ad2281f1e80b7c0e265f3b7138dee17fb2d160487604eee66
• Conversion source: scripts/convert_lesion_seg_model.py

scripts/convert_lesion_seg_model.py reconstructs the MONAI 3D UNet from the upstream config (channels=[32,64,128,256,320,320], strides=[2,2,2,2,2], PRELU, INSTANCE, num_res_units=1), loads the Safetensors weights, traces a static 1 x 1 x 128 x 128 x 128 tensor, exports opset 17 ONNX with weights inlined in a single file, then checks PyTorch vs ONNX Runtime output with a max-absolute-difference gate of < 1e-2.

The upstream model card describes 192³ patches, 1 mm spacing, one T1w input channel, background/stroke output classes, and optional TTA. CALMaR uses the same baseline weights but runs browser-bounded 128³ sliding-window patches with overlap=0.25, testTimeAugmentation=false, probability threshold 0.4, and minimum connected component size 30. At runtime the worker collapses the 2-channel [background, stroke] logits into a single stroke log-odds map (stroke - background) before sigmoid thresholding.

Brain extraction model provenance

CALMaR uses exactly one SynthStrip model asset for brain extraction:

• Manifest id: lnm-synthstrip
• Runtime filename: models/synthstrip.onnx
• Source version: SynthStrip-Hoopes2022-Apache-2.0
• Browser format: onnx-fp32
• Cache key: lnm-synthstrip-v1
• Size: 10294211 bytes
• SHA-256: 7b8eeecf3793a6c4510b9f5270ecc03d9c3262d26e08d568203a651ab4b84074
• Hosted source: sbollmann/lnm-webapp-models/models/synthstrip.onnx

The ONNX asset was converted from FreeSurfer's official main SynthStrip model v1 checkpoint:

• Original model: synthstrip.1.pt (main SynthStrip model v1, not synthstrip.nocsf.1.pt or the pediatric model)
• Original source: surfer.nmr.mgh.harvard.edu/pub/dist/freesurfer/synthstrip/models/synthstrip.1.pt
• Original SHA-256: 37417f802196186441aae3e7f385d94f8a98c64a88acaeaa2723af995c653e33
• Conversion source: neurodesk/vesselboost-webapp:scripts/convert_synthstrip.py

The app resolves this entry from web/models/manifest.json. runBrainExtraction() passes it to the module worker with the lnm-synthstrip model asset id. The worker runs web/js/modules/brain-extraction.js with the browser fast path (fast: true), WASM execution, no final dilation, SDT < 1 thresholding, largest connected component cleanup, and interior fill. The JS wrapper is the browser port of the FreeSurfer SynthStrip preprocessing and postprocessing path used by neurodesk/vesselboost-webapp.

Local development

npm install
bash web/setup.sh   # downloads ONNX Runtime WASM
bash web/run.sh     # serves http://localhost:8080/
npm test            # full Node-only suite: lint, manifest, overlap,
                    # function profiles, spatial checks, inference helpers,
                    # worker, app, html

Browser smoke tests

Optional. Run the Phase 1 manual-mask Yeo flow + the Phase 2a.1 auto-fired SynthStrip flow in headless Chromium. Not in npm test; requires a one-off browser install:

npx playwright install chromium
npm run test:smoke               # ~10 s on M-series; needs HF access

Node-side parity tests

Drive each ONNX pipeline directly via onnxruntime-node (no browser), against a real MNI152 anatomical T1. Pipeline-correctness checks (plausibility, not Dice — see commit notes).

npm run test:synthstrip-parity     # SynthStrip:        ~5 s
npm run test:lesion-seg-parity     # Lesion seg:        ~5 s; Dice >= 0.50 vs ds004884 ground truth
npm run test:registration-parity   # SynthMorph:       ~3 s (CPU EP); browser graph self-pair near-identity
npm run test:fc-weighted-sum-parity # FC weighted sum: ~1 s; identity case bit-exact

All fetch their respective ONNX models live from Hugging Face on first run (cached under web/models/_dev_cache/, gitignored). The lesion-segmentation parity uses one chronic-stroke subject from OpenNeuro ds004884 (Aphasia Recovery Cohort, Roth et al. 2024 — CC0); see tests/fixtures/ds004884-mini/SOURCE.md for attribution. Observed Dice on sub-M2051 ses-284: 0.5325.

License

TBD; intended to be open-source. Third-party assets retain their own licenses.