Every structure.
Its own object.

One subject.
One coordinate space.

Fusion is a local-first neuroimaging workspace. Give it structural scans for one subject, a T1 MRI plus any of T2, FLAIR or CT, and it registers every modality into one coordinate space, segments the brain into 133 labelled anatomical structures, builds a mesh per structure, and fuses it all into one linked workspace: 3D anatomy, slice review, per-structure metrics and exports.

It is research software, built for exploration, measurement and QC, not for diagnosis or surgical decisions. Segmentation output is prototype quality; every structure carries QC flags and a manual review workflow exists precisely because the software can be wrong.

One key.
Paste once.

First launch asks for a licence key (PFL1.…), shown on the download page immediately after checkout. The app verifies the signature offline and stores it locally; you won’t be asked again on that machine. The key covers the current billing period plus a grace window. After each renewal, revisit your download page for a fresh key.

One subject
at a time.

Choose a file per modality, press Ingest on each, then run the pipeline. T1 is the reference anatomy and required; every other modality is registered into T1 space. DICOM uploads are de-identified before storage. Patient identifiers are stripped and a de-id record is kept per subject.

No scan handy? Load synthetic phantom builds a clearly-labelled test subject, and a real demo subject is seeded on first launch. With several subjects pending, the Subjects panel offers Process all, a one-at-a-time batch queue.

Ingest → register
→ segment → measure.

Each stage shows as it runs: ingest check, registration into T1, UNesT segmentation, per-structure QC, mesh extraction, workspace build. Roughly 12–17 minutes on Apple Silicon for a T1+T2 subject; longer on CPU. GPU order is CUDA → Apple MPS → CPU, with automatic CPU fallback on out-of-memory. You can leave and return. The job continues on the local service.

Separate the brain.
Keep its relationships.

Structure tree (left): the full parcellation grouped by system. Toggle visibility per structure or group, isolate a subtree, or click to select and centre. Hidden by default: the whole-brain envelope, bone, white-matter interiors, and small off-brain fragments, which are preserved in data as a real QC finding.

Stage (centre): the 3D anatomy. Drag to rotate; Explode in five modes (all, hemisphere, deep, ventricles, selected) on a 0–100% slider; Labels pin names to structures; Slices renders the real T1 on orthogonal planes; Clip cuts on any axis in millimetres; modality chips overlay registered T2, FLAIR or CT at adjustable opacity.

Slice review (bottom): linked 2D slices in overlay, side, diff and checkerboard modes for judging registration quality. Clicking a 2D location marks it in 3D; selecting a structure centres the slices on it.

Floating panels: the Inspector (volume, laterality, hierarchy, QC flags, stability, repeatability, review verdicts), Registration QC (a verdict per registration), and Provenance (every pipeline step that touched the subject).

Fast hands.

Yours to take.

Trust is earned
per structure.

QC is part of the output, not an error state. Structures carry flags; flagged counts and stability scores show on subject cards and the stage header. Manual verdicts (PASS / QUESTIONABLE / FAIL plus notes) are recorded in the Inspector and Registration QC panel, stored with the subject, and appear in the report. A subject with unresolved verdicts is marked for review.

Your disk.
Your data.

Delete + clear removes every file and job record for a subject. No hidden copy. Uninstalling leaves the data directory behind; delete it manually for a full wipe.

When it argues.