Image Reconstruction Via Autoencoding Sequential Deep Image Prior
Authors: Ismail Alkhouri, Shijun Liang, Evan Bell, Qing Qu, Rongrong Wang, Saiprasad Ravishankar
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments, we validate the effectiveness of our method in various image reconstruction tasks, such as MRI and CT reconstruction, as well as in image restoration tasks like image denoising, inpainting, and non-linear deblurring. Our code is available at the Git Hub repository a Seq DIP. |
| Researcher Affiliation | Academia | Ismail R. Alkhouri 1,2, Shijun Liang 3, Evan Bell1, Qing Qu2, Rongrong Wang1,4, Saiprasad Ravishankar1,3 1Department of Computational Mathematics, Science, & Engineering, Michigan State University 2Department of Electrical Engineering & Computer Science, University of Michigan Ann Arbor 3Department of Biomedical Engineering, Michigan State University 4Department of Mathematics, Michigan State University |
| Pseudocode | Yes | Algorithm 1 Autoencoding Sequential Deep Image Prior (a Seq DIP). Input: Measurements y, forward operator A, number of input updates K, number of gradient updates N per input update, regularization parameter λ, and learning rate β. Output: Reconstructed image ˆx. |
| Open Source Code | Yes | Our code is available at the Git Hub repository a Seq DIP. |
| Open Datasets | Yes | For MRI, we use the fast MRI dataset3. The forward model is y Ax . The multi-coil data is obtained using 15 coils and is cropped to a resolution of 320 320 pixels. [...] For CT, we use the AAPM dataset4. [...] For the tasks of denoising, in-painting, and non-linear deblurring, we use the CBSD68 dataset5. |
| Dataset Splits | No | The paper describes using existing datasets (fast MRI, AAPM, CBSD68) and notes that their method does not require training data. For the supervised baseline (Var Net) it states "trained on 8000 data points from fast MRI" and "trained on 3000 data points from fast MRI", but it does not specify train/validation/test splits for their own experiments or for the general use of these datasets in the context of their method. |
| Hardware Specification | Yes | All the experiments are conducted on a single RTX5000 GPU machine. |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' and 'BART toolbox' but does not provide specific version numbers for these or any other software dependencies needed to replicate the experiments. |
| Experiment Setup | Yes | For the proposed a Seq DIP method in Algorithm 1, we use the Adam optimizer with learning rate of β = 0.0001. Furthermore, the regularization parameter is set to λ = 1 following the ablation study in Appendix C.4. We select N = 2 and K = 2000 following the ablation study in Appendix C.5. |