For training medical image registration models
OFG is a training framework that successfully unites learning-based methods with optimization techniques to enhance the training of learning-based registration models. OFG provides guidance with pseudo ground truth to the model by optimizing the model's output on-the-fly, which allows the model to learn from the optimization process and improve its performance.
![ofg_arch](https://private-user-images.githubusercontent.com/89094576/351052621-941c01da-c483-44c5-96b1-f5d9614f3100.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.rT7MucAoQiuze0K2ti9INgSbaGWQXbjfiuG9lsLt0aU)
OFG is a two stage training method, integrating optimization-based methods with registration models. It optimize the model's output in training time, this process generates a pseudo label on-the-fly, which will provide supervision for the model, yielding a model with better registration performance.
![benchmark](https://private-user-images.githubusercontent.com/89094576/351052852-7975ee17-57f9-40e8-9c21-d90addd60870.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.lO2Hl6MR3LuhCW5azY8kyZWhf4YOclZNa1S_NptkZQU)
OFG consistently improves the registration methods it is used on, and achieves state-of-the-art performance. It has better trainability than unsupervised methods while not using any manually added labels.
![ofg_lpba](https://private-user-images.githubusercontent.com/89094576/351053096-9f651a5d-3dfd-44c7-99a8-f304575bca5f.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.8T7sw6dUbT_9zHykPMZ9vkCzbrb7Rh-8krOgv3c6-Yk)
OFG provides much smoother deformation while also improving DSC of registration, combining into better overall registration performance across a wide range of modalities and datasets.
Cite our work when comparing results:
@article{ofg2024,
title={On-the-Fly Guidance Training for Medical Image Registration},
author={Yuelin Xin and Yicheng Chen and Shengxiang Ji and Kun Han and Xiaohui Xie},
year={2024},
eprint={2308.15216},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2308.15216},
}