FIRE：基于深度网络的无监督双向跨模态配准
计算机视觉与模式识别
2024-02-13 v1 图像与视频处理
摘要
跨模态图像配准是常规临床路径中许多应用的关键预处理步骤。本文提出一种无监督深度跨模态配准网络，可同时学习最优的仿射与非刚性变换。逆一致性是近期基于深度学习的跨模态配准算法普遍忽略的重要性质。我们通过所提出的多任务架构与新型综合变换网络来解决此问题。具体而言，该模型学习模态无关的潜在表示以执行循环一致的跨模态合成，并利用逆一致性损失学习一对变换将合成图像与目标图像对齐。因其结构形状，我们将该框架命名为FIRE。在多个序列脑MR数据以及模态内4D心脏Cine-MR数据的实验中，我们的方法表现出与流行基线方法相当或更优的性能。
引用
@article{arxiv.1907.05062,
title = {FIRE: Unsupervised bi-directional inter-modality registration using deep networks},
author = {Chengjia Wang and Giorgos Papanastasiou and Agisilaos Chartsias and Grzegorz Jacenkow and Sotirios A. Tsaftaris and Heye Zhang},
journal= {arXiv preprint arXiv:1907.05062},
year = {2024}
}
备注
We submitted this paper to a top medical imaging conference, srebuttal responded by the meta-reviewer. We were told that this work is not important and will not have big impact as the "reviewers were not enthusiastic". Here I publish the paper online for an open discussion. I will publish the code, the pre-trained model, the results, especially the reviews, and the meta reviews on github