Fourier-Net: Fast Image Registration with Band-Limited Deformation
Authors: Xi Jia, Joseph Bartlett, Wei Chen, Siyang Song, Tianyang Zhang, Xinxing Cheng, Wenqi Lu, Zhaowen Qiu, Jinming Duan
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Fourier-Net is then evaluated on two public 3D brain datasets against various state-of-the-art approaches. For example, when compared to a recent transformer-based method, named Trans Morph, our Fourier Net, which only uses 2.2% of its parameters and 6.66% of the multiply-add operations, achieves a 0.5% higher Dice score and an 11.48 times faster inference speed. |
| Researcher Affiliation | Academia | Xi Jia1, Joseph Bartlett1,2, Wei Chen1, Siyang Song3, Tianyang Zhang1, Xinxing Cheng1, Wenqi Lu4, Zhaowen Qiu5 , Jinming Duan1,6 1 School of Computer Science, University of Birmingham, UK 2 Department of Biomedical Engineering, University of Melbourne, Australia 3 Department of Computer Science and Technology, University of Cambridge, UK 4 Department of Computer Science, University of Warwick, UK 5 Institute of Information Computer Engineering, Northeast Forestry University, China 6 Alan Turing Institute, UK |
| Pseudocode | No | The paper describes the methodology using text and diagrams (Figure 2, Figure 3) but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/xi-jia/Fourier-Net. |
| Open Datasets | Yes | OASIS-1 dataset (Marcus et al. 2007) consists of a crosssectional collection of T1-weighted brain MRI scans from 416 subjects. In experiments, we use the pre-processed OASIS data1 provided by (Hoopes et al. 2021) to perform subject-to-subject brain registration, in which all 414 MRI scans are bias-corrected, skull-stripped, aligned, and cropped to the size of 160 192 224. |
| Dataset Splits | Yes | We randomly split this 2D dataset into 201, 12, and 201 images for training, validation, and test. After pairing, we end up with 40200 (201 200), 22 ([12-1] 2), and 400 ([201-1] 2) image pairs for training, validation, and test, respectively. |
| Hardware Specification | Yes | All deep models are trained with an Nvidia A100 GPU. |
| Software Dependencies | No | We implement our Fourier-Net using Py Torch, where training is optimized using Adam with a fixed learning rate of 0.0001. (PyTorch version not specified) |
| Experiment Setup | Yes | We implement our Fourier-Net using Py Torch, where training is optimized using Adam with a fixed learning rate of 0.0001. We tune built-in hyper-parameters on a held-out validation set. Specifically, we use MSE to train both 2D and 3D OASIS for 10 and 1000 epochs, respectively, and λ in L(Θ) is set to 0.01. For 3D OASIS, an additional Dice loss is used with its weight being set to 1. On 3D IXI, we train the models with NCC loss for 1000 epochs with λ = 5. |