We utilized Restormer (https://github.com/swz30/Restormer) as the backbone for our single-channel network and employed both Restormer and MoDL (https://github.com/hkaggarwal/modl) as backbones for our multi-channel network.
Overview of Latent Transformer
We introduced a 'Latent Transformer' featuring an encoder-decoder architecture with shared blocks, designed to model time-dependent relationships.
Each block of the Latent Transformer harnesses multi-layer and multi-head self-attention mechanisms to update the latent code through a weighted linear combination of itself and the latent codes from other time-frames, executed in a pixel-wise fashion.
@inproceedings{tanzer2023t1,
title={T1/T2 Relaxation Temporal Modelling from Accelerated Acquisitions Using a Latent Transformer},
author={T{\"a}nzer, Michael and Wang, Fanwen and Qiao, Mengyun and Bai, Wenjia and Rueckert, Daniel and Yang, Guang and Nielles-Vallespin, Sonia},
booktitle={International Workshop on Statistical Atlases and Computational Models of the Heart},
pages={293--302},
year={2023},
organization={Springer}
}