Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
X-Field: A Physically Informed Representation for 3D X-ray Reconstruction
Authors: Feiran Wang, Jiachen Tao, Junyi Wu, Haoxuan Wang, Bin Duan, Kai Wang, Zongxin Yang, Yan Yan
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments show that X-Field achieves superior visual fidelity on both real-world human organ and synthetic object datasets, outperforming state-of-the-art methods in X-ray NVS and CT Reconstruction. Our code is available on the project page: https://github.com/Brack-Wang/X-Field. We evaluate X-Field on X-ray reconstruction across various modalities, including real-world human organ datasets and simulated object datasets. Furthermore, we compare X-Field with existing methods in sparse-view CT reconstruction based on Novel View Synthesis (NVS) results. Experiments demonstrate that X-Field outperforms state-of-the-art (SOTA) methods in both NVS and CT reconstruction. |
| Researcher Affiliation | Academia | 1University of Illinois Chicago 2University of Michigan 3National University of Singapore 4Harvard Medical School |
| Pseudocode | Yes | Algorithm 1 Segment Length Correction with Intersections Input: (z0, z1, . . . , zn 1): Sorted depths (l0, l1, . . . , ln 1): Initial segment lengths Output: Updated segment lengths l0, l1, . . . , ln 1 1: for i = 0 to n 1 do 2: if i == 0 then 3: l0 l0 4: z z0, l l0 5: else 6: if zi < z + 1 7: li max(0, (zi + 1 2li) (z + 1 2l)) 8: else 9: li min(li, ( 1 2li + zi) (z + 1 2l)) 10: end if 11: if li = 0 then 12: [zi + 1 2li li, zi + 1 2li] ; z zi, l li 13: end if 14: end if 15: end for |
| Open Source Code | Yes | Our code is available on the project page: https://github.com/Brack-Wang/X-Field. |
| Open Datasets | Yes | We conduct experiments on a large-scale X3D dataset containing 15 scenes with two collections: Human Organs, derived from real-world medical datasets; and Daily Objects, generated from synthetic datasets. Specifically, chest scans are sourced from LIDC-IDRI [49], pancreas CT scans from Pancreas-CT [50]. The remaining objects are obtained from VOLVIS [51] and the open scientific visualization dataset [52]. We evaluate our method on the FIPS dataset [61], which provides real 2D X-ray projections. |
| Dataset Splits | Yes | For highly sparse-view novel view synthesis, 5 and 10 views are used for training and 50 views are used for testing. To further assess model performance and scalability, we generate 50, 25, and 15 views for evaluating the performance under sparse-view synthesis. For training, we uniformly capture 5, 10, 15, 25, and 50 views within 0 to 180 . For testing, we randomly capture 50 views in the same range. |
| Hardware Specification | Yes | All experiments were conducted on a single RTX 6000 Ada GPU. |
| Software Dependencies | No | We adopt the tomography toolbox TIGRE [53] to capture projections from CT volumes in the range of 0 180 with minor scatter and electronic noise. |
| Experiment Setup | Yes | In our experiments, the position learning rate was set to an initial and final value of 0.0002, while the density and scaling learning rates were both initialized and maintained at 0.01 and 0.005, respectively. The structural dissimilarity loss weight (λDSSIM) was set to 0.25. To refine the density of kernels, a minimum density threshold of 1 10 5 was applied. The total number of kernels was capped at 500,000 to balance computational efficiency with accuracy. |