How rohankumar-1/brain-model works
BrainQuery is a 3D brain CT/MRI foundation-model package whose main workflow turns raw cohort data into registered, indexed volumes, trains self-supervised and task-conditioned models, then evaluates them on held-out tasks and a versioned benchmark. Its other documented capabilities include few-shot task learning, frozen-encoder transfer, segmentation, and published-encoder comparisons. The graph emphasizes those domain stages and the shared encoder/data layers. Some subsystems are represented by their exact enclosing directories because their implementation was not sampled; edges reflect README statements or shown source references.
Main entry points: store.py Processed store · volumes.py SSL volume data · model.py BrainQuery model · downstream Frozen transfer
Components 24
Preprocessing
- Dataset fetch
brain/datasets - Source readers
brain/prep/sources.py - Volume processing
brain/prep/pipeline.py - Processed store
brain/prep/store.py - Volume index
brain/prep/index.py
Model training
- SSL volume data
brain/data/volumes.py - LeJEPA training
brain/train/lejepa3d.py - Joint training
brain/train/joint.py - Training engine
brain/train/engine.py
Task learning
- Task registry
brain/multitask/registry.py - Label building
brain/multitask/labels.py - Task batches
brain/multitask/data.py - BrainQuery model
brain/multitask/model.py - Multitask training
brain/train/multitask.py - Few-shot learning
brain/multitask/fewshot.py
Evaluation
- Frozen transfer
brain/downstream - Published adapters
brain/downstream/adapters - Benchmark suite
brain/bench - Head evaluation
brain/multitask/evaluate.py - Evaluation reports
brain/eval/report.py
Shared core
- 3D ViT encoder
brain/models/vit3d.py - Intensity windowing
brain/common/windowing.py
Other components
- Researcher
- Raw cohorts
Connections 31
- Researcher → Dataset fetch: fetches data
README.md - Raw cohorts → Dataset fetch: supplies archives
README.md - Dataset fetch → Source readers: provides inputs
README.md - Source readers → Volume processing: reads cases
brain/prep/pipeline.py - Volume processing → Processed store: writes volumes
README.md - Processed store → Volume index: indexes volumes
brain/prep/index.py - Volume index → SSL volume data: selects splits
README.md - Processed store → SSL volume data: loads volumes
brain/data/volumes.py - SSL volume data → LeJEPA training: provides crops
README.md - LeJEPA training → Training engine: runs training
brain/train/engine.py - LeJEPA training → 3D ViT encoder: trains encoder
README.md - Joint training → Training engine: runs training
README.md - Joint training → 3D ViT encoder: trains jointly
README.md - Task registry → Label building: defines tasks
brain/multitask/labels.py - Label building → Task batches: supplies targets
README.md - Processed store → Task batches: loads scans
brain/multitask/data.py - Task batches → Multitask training: provides batches
README.md - Multitask training → BrainQuery model: trains model
brain/train/multitask.py - BrainQuery model → 3D ViT encoder: uses encoder
README.md - Multitask training → Training engine: runs training
brain/train/engine.py - Researcher → Few-shot learning: learns task
README.md - Few-shot learning → BrainQuery model: adapts model
brain/multitask/fewshot.py - Researcher → Frozen transfer: runs transfer
README.md - Frozen transfer → 3D ViT encoder: embeds scans
brain/downstream/embed3d.py - Published adapters → Frozen transfer: provides adapters
README.md - Researcher → Benchmark suite: runs benchmark
README.md - Benchmark suite → Frozen transfer: evaluates encoders
README.md - Researcher → Head evaluation: scores heads
README.md - Head evaluation → BrainQuery model: evaluates heads
README.md - Benchmark suite → Evaluation reports: summarizes results
README.md - Intensity windowing → SSL volume data: windows intensities
brain/data/volumes.py