Ask for a pathway.
Get streamlines.
Pulling one white-matter pathway out of a whole-brain tractogram still means expert manual ROI placement or a hand-tuned, atlas-specific pipeline. BundleMind takes the question the way you'd say it and returns streamlines — anatomically grounded, on your own laptop.
16 questions, no manual ROIs.
One continuous pass over the same HCP842 whole-brain atlas, holding on each selection as it lands. Every one of them came from the plain-language query shown with it — click a marker, or any question below, to jump there.
Simple naming
Functional
Spatial
Compositional
Negation
Categorical
Clinical
Two query paths, one router.
A free-form query enters a router that tries the deterministic filter first and falls back to a learned embedding when the phrasing is open-ended or out of vocabulary. Both paths end at the same region-grounded executor.
Free-form query
Typed the way a neuroscientist, a student or a clinician would actually say it — no controlled vocabulary, no ROI painting.
Filter
A local language model translates the question into a compact JSON program. It decodes under a schema, so the output is always parseable.
Embedding
A learned text↔streamline space ranks atlas bundles by similarity, catching phrasings the filter path has no rule for.
Grounded executor
Cytoarchitectonic waypoint labels decide which streamlines actually belong — the same check whichever path produced the program.
Language in, a small program out.
One compact DSL sits between the language model and a deterministic geometric executor. It is the whole interface — everything the system can be asked is expressible in these terms.
connectspasses_throughendpoint_inandornotlength_betweenatlas_bundlebundle_typescore_above // “tracts through thalamus under 80 mm”
{"and": [
{"passes_through": "thalamus"},
{"length_between": {"max_mm": 80}}
]}Anatomy decides, not geometry alone.
A streamline joins a bundle only when its waypoint labels overlap that bundle's endpoint signature — two or more shared labels per hemisphere, with endpoint-based hemisphere guards. Cross-bundle continuations survive; spurious cortex-wide and wrong-hemisphere streamlines are dropped.
+52% overall F1.
Matched run on the same 88-query benchmark. Constraining generation to the schema and grounding region names before generation is what moves the numbers — most of the gain lands on the hardest categories.
Want to run BundleMind on your tractograms?
BundleMind is in active development at Indiana University. For access or to collaborate, reach out.
srchig@iu.edu