Autoregressive Image Diffusion: Generation of Image Sequence and Application in MRI
Authors: Guanxiong Luo, Shoujin Huang, Martin Uecker
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Models trained with fast MRI dataset are evaluated comprehensively. The results show that the AID model can robustly generate sequentially coherent image sequences. In MRI applications, the AID can outperform the standard diffusion model and reduce hallucinations, due to the learned inter-image dependencies. |
| Researcher Affiliation | Academia | Guanxiong Luo University Medical Center Göttingen guanxiong.luo@med.uni-goettingen.de Shoujin Huang Shenzhen Technology University Martin Uecker Graz University of Technology uecker@tugraz.at |
| Pseudocode | Yes | Algorithm 1 Sample the posterior in {p(xn|yn, x0 <n)|1 < n < N} using autoregressive diffusion model as prior. |
| Open Source Code | Yes | The project code is available at https://github.com/mrirecon/aid. |
| Open Datasets | Yes | The image space model was trained on brain images that are from the fast MRI training dataset, which includes T1-weighted (some with post-contrast), T2-weighted, and FLAIR images [27]. |
| Dataset Splits | Yes | The image space model was trained on brain images that are from the fast MRI training dataset, which includes T1-weighted (some with post-contrast), T2-weighted, and FLAIR images [27]. |
| Hardware Specification | Yes | All the training was performed on 4 NVIDIA A100 GPUs with 80GB memory. |
| Software Dependencies | No | The paper mentions software like Open AI's guided diffusion codebase and the einops library, but does not provide specific version numbers for these or other key software dependencies. |
| Experiment Setup | Yes | The models were trained using the Adam optimizer with a learning rate of 10^-4 and a batch size of 1 for image space model and 4 for latent space model. The length of conditioning sequence N for brain and cardiac models are 10 and 42. Setting parameters: T = 1000, λ = 1, K = 4 for Algorithm 1. |