TrackletGPT

B-Spline Streamline Tokenization for White Matter Segmentation

Organization: IIT Mandi

Key Result & Impact Benchmark

Set new benchmark accuracy for non-invasive white matter tract anatomical segmentation.

Technical Overview

TrackletGPT is a peer-reviewed research paper accepted at the IEEE International Symposium on Biomedical Imaging (ISBI 2026, London, UK) with Conference Rank A. Building upon TractoGPT, TrackletGPT introduces a novelty in 3D streamline tokenization: representing fiber tract sub-sequences as continuous B-Spline 'tracklets'. By encoding geometric curvature parameters into discrete tokens, the GPT transformer pretrains on tracklet completion and dual-masking, achieving state-of-the-art white matter tract segmentation accuracy across complex crossing-fiber brain anatomical regions.

Key Technical Highlights

Accepted at IEEE ISBI 2026 (London, UK) with Conference Rank A classification.
Formulated B-Spline continuous curve tokens to preserve spatial streamline trajectory continuity.
Pretrained GPT transformer architecture on dual-masked tracklet reconstruction tasks.

BibTeX Citation

@inproceedings{goel2026trackletgpt,
  title={TrackletGPT: A GPT architecture for White Matter Segmentation},
  author={Goel, Anoushkrit and Nigam, Aditya and Bhavsar, Arnav},
  booktitle={IEEE International Symposium on Biomedical Imaging (ISBI)},
  year={2026}
}

Technologies & Frameworks

3D Spatial AI

B-Spline Tokenization

GPT Transformers

PyTorch3D

Plotly