Noise2Inverse LPD enables self-supervised CT reconstruction without ground truth
New algorithm trains directly on noisy scans, outperforming supervised methods like U-Net.
X-ray computed tomography (CT) reconstruction is an ill-posed inverse problem, especially under low-dose or sparse-angle conditions. Traditional learned methods like Learned Primal-Dual (LPD) achieve high quality but require paired clean/noisy data for supervised training—a major bottleneck in clinical settings. Now, a team of researchers introduces Noise2Inverse Learned Primal-Dual (N2I-LPD), which extends the Noise2Inverse self-supervised framework to iterative reconstruction. The key insight: noise in CT measurements is statistically independent across different angular projections. By splitting measurements into two subsets, the algorithm can learn to reconstruct without ever seeing a clean reference.
Evaluated on simulated and real CT data, N2I-LPD consistently outperforms both classical methods (e.g., FBP) and U-Net-based approaches trained within the same Noise2Inverse framework. The method reconstructions show better preservation of fine details and reduced artifacts, even at extremely low photon counts (e.g., 10% of normal dose). This work demonstrates that combining iterative learned architectures with self-supervised denoising strategies can unlock practical CT imaging scenarios where ground-truth data is impossible to acquire—such as in living patients or industrial in-situ monitoring. The code is available on GitHub.
- N2I-LPD eliminates the need for ground-truth images by leveraging noise independence across CT angular rotations.
- It combines a learned iterative reconstruction operator (Learned Primal-Dual) with the Noise2Inverse self-supervised framework.
- Outperforms a U-Net trained under the same self-supervised setup, particularly in low-dose (10% photon count) scenarios.
Why It Matters
Enables high-quality CT imaging in low-dose settings where ground truth is unavailable, reducing patient radiation exposure.