Learning White Matter Streamline Representations Using Transformer-based Siamese Networks with Triplet Margin Loss
Shenjun Zhong, Zhaolin Chen, Gary F. Egan · Proceedings on CD-ROM - International Society for Magnetic Resonance in Medicine. Scientific Meeting and Exhibition/Proceedings of the International Society for Magnetic Resonance in Medicine, Scientific Meeting and Exhibition · 2023
Robust latent representation of white matter streamlines are critical for parcellating streamlines. This work introduced a novel transformer-based siamese network with triplet margin loss, that learns to embed any lengths of streamlines into fixed-length latent representations. Results showed that a minimum of two layers of transformer encoders were sufficient to model streamlines with a very limited number of training data.