TractoGPT
Dual-Masked GPT Architecture for White Matter Segmentation
Established state-of-the-art segmentation accuracy across challenging brain fiber crossing regions.
Technical Overview
TractoGPT is a peer-reviewed research paper accepted at the 23rd IEEE International Symposium on Biomedical Imaging (ISBI 2025, Houston, Texas). TractoGPT models 3D tractography streamlines as spatial point cloud tokens. The model introduces a dual-masking self-supervised pretraining objective: masking both individual point coordinates within a patch and whole spatial patches across the brain. Once pretrained on point cloud reconstruction, the decoder-only GPT transformer is fine-tuned for white matter tract classification, outperforming Sherbrooke Connectivity Imaging Lab (SCIL Canada)'s state-of-the-art model, FIESTA.
Key Technical Highlights
- Accepted at IEEE ISBI 2025 (Houston, Texas).
- Outperformed SCIL Canada's state-of-the-art FIESTA model on anatomical tract segmentation.
- Introduced dual-masking self-supervised pretraining tailored specifically for 3D spatial point sequences.
BibTeX Citation
@article{goel2024tractogpt,
title={TractoGPT: A GPT architecture for White Matter Segmentation},
author={Goel, Anoushkrit and Nigam, Aditya and Bhavsar, Arnav},
journal={arXiv preprint arXiv:2411.08187},
year={2024}
}