Unsupervised Self-Prior Embedding Neural Representation for Iterative Sparse-View CT Reconstruction
Authors: Xuanyu Tian, Lixuan Chen, Qing Wu, Chenhe Du, Jingjing Shi, Hongjiang Wei, Yuyao Zhang
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on multiple CT datasets show that our unsupervised Spener method achieves performance comparable to supervised state-of-the-art (SOTA) methods on in-domain data while outperforming them on out-of-domain datasets. Additionally, the paper includes sections such as "Experiments", "Experimental Settings", "Methods in Comparison & Metrics", "Comparison with SOTA Methods", and "Ablation Studies", all indicating empirical evaluation. |
| Researcher Affiliation | Academia | 1 School of Information Science and Technology, Shanghai Tech University, Shanghai 201210, China 2 Lingang Laboratory, Shanghai 200031, China 3 Electrical and Computer Engineering, University of Michigan, MI 48105, United States 4 Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology, Wuhan 430030, China 5School of Biomedical Engineering, Shanghai Jiao Tong University, Shanghai 200127, China EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes the proposed method in narrative text and through equations, but does not include a dedicated section for pseudocode or an algorithm block. Figure 1 provides an overview diagram but is not pseudocode. |
| Open Source Code | Yes | Code https://github.com/Meiji Tian/Spener |
| Open Datasets | Yes | We evaluate the proposed method on three public datasets: AAPM 2016 low-dose CT grand challenge (Mc Collough et al. 2017), COVID-19 (Shakouri et al. 2021) and CMB-CRC head dataset (Cancer Moonshot Biobank 2022). |
| Dataset Splits | Yes | The AAPM dataset comprises 5936 full-dose CT images from 10 patients with 1mm slice thickness. In the experiments, we specifically select 1600 slices from 8 patients as the training set, 200 slices from 1 patient as the validation set and 10 slices from 1 patient as the test set. |
| Hardware Specification | No | The paper does not explicitly mention any specific hardware details such as GPU models, CPU types, or memory configurations used for running the experiments. |
| Software Dependencies | No | The paper mentions using the 'torch-radon library in Python' and the 'classical BM3D algorithm' but does not provide specific version numbers for these software dependencies or any other key components. |
| Experiment Setup | Yes | In the Spener, we adopt the FBP reconstruction result as the initial prior image. The image encoder includes a two-layer CNN network, with a convolution kernel size of 3 and a feature dim of 48. For the hash encoding used in our model, we set its hyper-parameters as follows: base resolution Nmin = 2, maximal hash table size T = 224, and resolution growth rate b = 1.95. For the iteration process, the training epochs are configured as follows: 1000 epochs for t = 1 and 250 for the subsequent iterations. for the regularization, we utilize the BM3D denoiser with σ = 0.01 . Therefore, we set λ = 2.5 for the all experiments. |