autobidsify 0.6.0

Automated BIDS standardization tool powered by LLM-first architecture

autobidsify

Automated Brain Imaging Data Structure (BIDS) standardization tool powered by LLM-first architecture.

Features

• General compatibility: Handles diverse dataset structures (flat, hierarchical, multi-site)
• Multi-modal support: MRI, fNIRS, and mixed modality datasets
• Intelligent metadata extraction: Automatic participant demographics from DICOM headers, documents, and filenames
• Format conversion: DICOM→NIfTI, JNIfTI→NIfTI, .mat/.nirs→SNIRF, and more
• Multi-LLM support: OpenAI (gpt-4o, gpt-5.1, o1, o3) and Qwen (via Ollama or DashScope)
• Evidence-based reasoning: Confidence scoring and provenance tracking for all decisions

Supported Formats

Input formats:

• MRI: DICOM (.dcm), NIfTI (.nii, .nii.gz), JNIfTI (.jnii, .bnii)
• fNIRS: SNIRF (.snirf), Homer3 (.nirs), MATLAB (.mat)
• Documents: PDF, DOCX, TXT, Markdown

Output: Compliant to BIDS specification (v1.10.0)

Installation

pip install autobidsify

Optional dependencies:

# For DICOM conversion
apt-get install dcm2niix          # Ubuntu/Debian
brew install dcm2niix             # macOS

# For BIDS validation
npm install -g bids-validator

# For Qwen models (local)
# Install Ollama from https://ollama.com/download
ollama pull qwen2.5-coder:7b
pip install ollama

Set API key:

# OpenAI
export OPENAI_API_KEY="your-key-here"

# Qwen via DashScope (optional cloud alternative to Ollama)
export DASHSCOPE_API_KEY="your-key-here"

Quick Start

# Full pipeline (one command)
# With dataset description (recommended for better metadata extraction)
autobidsify full \
  --input /path/to/your/data \
  --output outputs/my_dataset \
  --model gpt-4o \
  --modality mri \
  --nsubjects 10 \
  --id-strategy numeric \
  --describe "Your dataset description here"

# Step-by-step execution
autobidsify ingest  --input data/ --output outputs/run
autobidsify evidence --output outputs/run --modality mri
autobidsify trio   --output outputs/run --model gpt-4o
autobidsify plan   --output outputs/run --model gpt-4o
autobidsify execute  --output outputs/run
autobidsify validate --output outputs/run

Command Options

--input PATH            Input data (archive or directory)
--output PATH           Output directory
--model MODEL           LLM model (default: gpt-4o)
--modality TYPE         Data modality: mri | nirs | mixed
--nsubjects N           Number of subjects (optional, auto-detected if omitted)
--describe "TEXT"       Dataset description (recommended for metadata accuracy)
--id-strategy STRATEGY  Subject ID strategy: auto | numeric | semantic (default: auto)

Supported Models

OpenAI:

--model gpt-4o           # Highly recommended, stable
--model gpt-4o-mini      # Faster, cheaper
--model gpt-5.1          # Not that ecommended, latest

Qwen (via Ollama, local):

--model qwen3-coder-next:latest     # Recommended
--model qwen3-coder-careful:latest  # Recommended
--model qwen2.5-coder:7b            # Slow and sometimes inaccurate, not recommended

Qwen (via rest-api):

export OLLAMA_BASE_URL=http://your-server.com:xxxx

Pipeline Stages

Stage	Command	Input	Output	Purpose
1	ingest	Raw data	ingest_info.json	Extract/reference data
2	evidence	All files	evidence_bundle.json	Analyze structure, detect subjects
3	classify	Mixed data	classification_plan.json	Separate MRI/fNIRS (optional)
4	trio	Evidence	BIDS trio files	Generate metadata files
5	plan	Evidence + trio	BIDSPlan.yaml	Create conversion strategy
6	execute	Plan	bids_compatible/	Execute conversions
7	validate	BIDS dataset	Validation report	Check compliance

Output Structure

outputs/my_dataset/
├── bids_compatible/              # Final BIDS dataset
│   ├── dataset_description.json
│   ├── README.md
│   ├── participants.tsv
│   ├── sub-001/
│   │   ├── anat/
│   │   │   └── sub-001_T1w.nii.gz
│   │   └── func/
│   │       └── sub-001_task-rest_bold.nii.gz
│   └── derivatives/              # Unprocessed files (original structure)
│       └── sub-001/
│           └── ...
└── _staging/                     # Intermediate files
    ├── evidence_bundle.json
    ├── BIDSPlan.yaml
    └── conversion_log.json

Examples

Example 1: Single-site MRI study

autobidsify full \
  --input brain_scans/ \
  --output outputs/study1 \
  --nsubjects 50 \
  --model gpt-4o \
  --modality mri

Example 2: Multi-site dataset with description

autobidsify full \
  --input camcan_data/ \
  --output outputs/camcan \
  --model gpt-4o \
  --modality mri \
  --id-strategy semantic \
  --describe "Cambridge Centre for Ageing and Neuroscience: 650 participants, ages 18-88, multi-site MRI study"

Example 3: fNIRS dataset

autobidsify full \
  --input fnirs_study/ \
  --output outputs/fnirs \
  --model gpt-4o \
  --modality nirs \
  --describe "Prefrontal cortex activation during cognitive tasks, 30 subjects"

Example 4: Using Qwen (local, no API cost)

ollama serve
autobidsify full \
  --input data/ \
  --output outputs/run \
  --model qwen2.5-coder:7b \
  --modality mri

Architecture

LLM-First Design:

• Python: Deterministic operations — file I/O, regex-based subject detection, format conversion, BIDS validation
• LLM: Semantic understanding — dataset description, metadata extraction, scan type classification, license normalization
• Hybrid: Python analyzes ALL files for completeness; LLM sees representative samples for semantic decisions

Requirements

• Python
• OpenAI API key (or Ollama for local Qwen models)
• dcm2niix for DICOM conversion
• bids-validator for validation

Current Status

Version: 0.6.0

Tested datasets:

• Visible Human Project (flat structure, DICOM CT)
• CamCAN (hierarchical, multi-site, 30+ subjects)
• 1-FRESH-Motor (fNIRS, existing BIDS format)
• fNIRS tinnitus dataset (flat structure, .nirs files)

Known limitations:

• Mixed modality classification (Stage 3) is experimental
• .mat fNIRS conversion assumes Homer3-compatible variable naming

Contributing

We need YOUR datasets to improve robustness. Please test and report issues at: https://github.com/cotilab/autobidsify/issues