Tract-RLformer
Tract-Specific Reinforcement Learning & Transformer Tracking
Markedly improved fiber tracking accuracy and cross-dataset generalizability across clinical dMRI scans.
Technical Overview
Tract-RLformer is a peer-reviewed research paper published at the 27th International Conference on Pattern Recognition (ICPR 2024, Kolkata). The network utilizes a hybrid two-stage training strategy combining supervised pretraining with reinforcement learning policy refinement. By training tract-specific agent policies within an OpenAI Gym environment over dMRI vector fields, Tract-RLformer directly tracks and delineates white matter bundles of interest without requiring post-hoc anatomical filtering. It markedly outperformed established tracking algorithms from Sherbrooke Connectivity Imaging Lab (SCIL Canada) including Track2Learn, DeepTract, and Particle Filtering Tracking (PFT).
Key Technical Highlights
- Published in Springer ICPR 2024 conference proceedings.
- Outperformed SCIL Canada's benchmark tracking algorithms (Track2Learn, DeepTract, PFT).
- Eliminated post-processing segmentation steps by learning direct goal-directed RL fiber tracking policies.
BibTeX Citation
@inproceedings{goel2024tract,
title={Tract-RLFormer: A Tract-Specific RL Policy Based Decoder-Only Transformer Network},
author={Goel, Anoushkrit and Nigam, Aditya and Bhavsar, Arnav},
booktitle={International Conference on Pattern Recognition},
pages={279--295},
year={2024},
organization={Springer}
}