Register histological brain sections to BrainGlobe atlases.
A VLM estimates position, an image-gen model lays down atlas colors, Elastix warps the rest.

How it works

A VLM agent (default langslice-gemma-4-E4B) inspects the slice, explores candidate atlas planes through tool calls, and submits an AP coordinate. Image generation then produces an atlas-colored target from the histology, and itk-elastix recovers a dense B-spline deformation. Results export to VisuAlign-compatible JSON for QUINT / ABBA.

Quick start

conda env create -f environment.yml
conda activate langslice
pip install -e .
cp .env.example .env  # add AI Studio / Vertex / OpenAI keys

# Position estimation
langslice estimate slice.png

# End-to-end registration at a known atlas position
langslice register slice.png --position 3.9

Full CLI: langslice --help. Pipeline detail: docs/index.md.

Model Hub Transition

LangSlice is migrating training/data helpers into shared packages under models/:

• models/langslice-traces/langslice_traces
• models/synthdata/synthdata
• models/training-core/langslice_training
• models/data/langslice_data

Training entrypoints are exposed as small launchers: langslice-single-turn-rl, langslice-isft, and langslice-sft-train.

Public model-card metadata for the released variant is in models/langslice-gemma-4/variants/langslice-gemma-4-e4b/README.md.

Links

• langslice-gemma-4-E4B — the v1.0 fine-tune
• SliceBench — self-contained position-estimation benchmark
• tauri-gui/ — desktop app
• docs/ — full pipeline + harness internals