Design Choices in Splitting-Based Self-Supervised Sparse-View CT Reconstruction

Self-supervised learning has become a standard approach for sparse-view CT reconstruction, allowing models to learn from incomplete measurements without requiring fully sampled ground-truth images. However, the field has lacked a unified understanding of how different design choices—such as how data is partitioned, how missing data is preprocessed, and how the final image is inferred—affect reconstruction quality. This paper introduces a unified framework to decompose these components and systematically benchmarks them across different noise regimes.

The authors decompose splitting-based reconstruction into three distinct stages: partitioning (how the sinogram is divided), preprocessing (how masked data is filled before reconstruction), and inference (how sub-reconstructions are aggregated). They test various strategies, including lattice-based and angular masking, and introduce a new 'multi-partition' approach that enforces consistency across multiple geometric configurations simultaneously. The methods are evaluated using both synthetic LoDoPaB-CT data (with independent and correlated noise) and the real-world 2DeteCT dataset, using both standard metrics (PSNR, SSIM) and perceptual metrics (LPIPS, HaarPSI).

The study reveals that the choice of partitioning is not universal but highly dependent on the noise characteristics of the imaging system. Lattice-based splitting, which samples pixels in a grid, performs best when noise is independent. In contrast, angular masking, which removes entire projection angles, is significantly more robust when dealing with correlated noise, such as that caused by scintillator blur in real CT detectors. Furthermore, the authors demonstrate that multi-partition splitting consistently outperforms standard single-partition methods, and that perceptual metrics are more effective than PSNR at capturing subtle, diagnostically relevant reconstruction artifacts.