TractoInferno

TractoInferno provides a multi-site diffusion MRI (dMRI) tractography database to develop and benchmark machine-learning algorithms for dMRI-based fiber tracking.

Key Features:

• Extensive Data Collection: 284 samples acquired from six different sites using 3T MRI scanners, including research-oriented and clinical-like human acquisitions.
• Comprehensive Data Types: Includes T1-weighted images, single-shell dMRI acquisitions, spherical harmonics fitted to the dMRI signal, Fiber Orientation Distributions (ODFs), and reference streamlines delineating 30 distinct fiber bundles generated using four different tractography algorithms.
• Reference Streamlines and Masks: Provides reference streamlines for 30 fiber bundles and the masks necessary for running tractography algorithms.
• Quality Control: Rigorous manual quality control performed at multiple stages of data processing.
• Resource Requirements: Development involved approximately 20,000 CPU-hours, 200 man-hours for manual QC, 3,000 GPU-hours for training baseline models, ~4 terabytes of intermediate storage, and a final database size of 350 gigabytes.
• Standardized Evaluation Protocol: Supplies a standardized dataset and evaluation protocol for machine-learning tractography benchmarking.

Scientific Applications:

• Algorithm Benchmarking: Used to benchmark the learn2track algorithm and five variations of a recurrent neural network architecture.
• Development and Evaluation of ML Tractography: Enables development, testing, and comparison of machine-learning algorithms aimed at improving dMRI-based fiber tracking.
• Research and Clinical Studies: Supports basic neuroscience research and clinical-like applications by providing both research-oriented and clinical-like acquisitions.

Methodology:

Data were acquired across six sites on 3T MRI scanners; spherical harmonics were fitted to the dMRI signal and Fiber Orientation Distributions (ODFs) computed; reference streamlines delineating 30 fiber bundles were generated using four tractography algorithms; masks were provided and rigorous manual quality control was applied; baseline models were trained using GPU resources.