Overview
This work examines the influence of design choices in splitting-based self-supervised sparse-view Computed Tomography (CT) reconstruction. The research introduces a unified framework to analyze and compare existing methods by decomposing the reconstruction process into three components: partitioning strategy, preprocessing, and inference. The study also explores two incremental extensions: multi-partition splitting and an alternative inference strategy within this framework.
Research Context
Self-supervised data splitting has emerged as a method for sparse-view CT reconstruction. This paradigm allows for training from incomplete measurements without requiring fully sampled ground truth data. Despite its promise, the impact of specific design choices, including the partitioning strategy, preprocessing steps, and inference mechanisms, remains underexplored.
Approach
The researchers developed a unified framework to systematically evaluate splitting-based reconstruction methods. This framework allows for controlled comparison by isolating and varying the partitioning strategy, preprocessing techniques, and inference procedures. Two incremental extensions were also integrated for evaluation: multi-partition splitting and an alternative inference strategy.
Experiments were conducted on simulated LoDoPaB-CT data, incorporating both independent and correlated noise conditions. Validation was performed using the real-world 2DeteCT dataset.
The evaluation utilized complementary perceptual and structural metrics in addition to standard quantitative measures. These included LPIPS (Learned Perceptual Image Patch Similarity) and HaarPSI (Haar wavelet-based Perceptual Similarity Index), alongside PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity Index Measure).
Findings
- The optimal partitioning strategy is strongly dependent on the structure of the measurement noise.
- Lattice-based splitting demonstrates favorable performance when independent noise is present.
- Angular masking exhibits greater robustness under conditions of correlated noise and when applied to real measured data.
- Multi-partition splitting consistently improves upon pure projection-wise splitting in several experimental settings.
- Perceptual and structural metrics, specifically LPIPS and HaarPSI, revealed differences between masking strategies that were less apparent when relying solely on PSNR and SSIM.
Why This Matters
These findings offer practical guidelines for the design of self-supervised sparse-view CT reconstruction methods. The research highlights the limitations associated with common independence assumptions in realistic imaging environments and provides insights into selecting appropriate strategies based on noise characteristics and data types.
Key Limitations Mentioned by Researchers
The study implicitly highlights limitations related to common independence assumptions prevalent in current designs for realistic imaging environments, suggesting that these assumptions may not hold universally.