TractoEmbed
Modular Multi-Embedding Framework for White Matter Segmentation
Achieved superior F1-score and Dice overlap coefficients on HCP & TRACULA anatomical tract datasets.
Technical Overview
TractoEmbed is a peer-reviewed research paper published at the 27th International Conference on Pattern Recognition (ICPR 2024, Kolkata). The paper proposes a modular multi-embedding architecture that extracts representations across three hierarchical levels: individual streamline level (via 1D CNN), regional patch level (via discrete Variational Autoencoder dVAE), and global bundle level (via PointNet). By fusing these multi-scale spatial embeddings, TractoEmbed classifies streamlines into 72 anatomical white matter tracts, outperforming Harvard Medical School's benchmark TractCloud model.
Key Technical Highlights
- Published in Springer ICPR 2024 proceedings.
- Outperformed Harvard Medical School's TractCloud benchmark on white matter tract segmentation.
- Combined 1D CNN, dVAE, and PointNet architectures into a unified hierarchical spatial feature extractor.
BibTeX Citation
@inproceedings{goel2024tractoembed,
title={TractoEmbed: A Modular Multi-embedding Framework for White Matter Segmentation},
author={Goel, Anoushkrit and Nigam, Aditya and Bhavsar, Arnav},
booktitle={International Conference on Pattern Recognition},
pages={262--278},
year={2024},
organization={Springer}
}