CTO neural operator reconstructs sparse-view CT across resolutions, 500x faster
One CT model handles any sampling rate with 3.4dB better PSNR than CNNs
Researchers from Caltech, led by Aujasvit Datta and Anima Anandkumar, have released Computed Tomography neural Operator (CTO), the first neural operator (NO) framework for sparse-view CT reconstruction. Sparse-view CT reduces radiation and scan time by reconstructing images from fewer X-ray projections, but standard deep learning models like CNNs overfit to fixed acquisition settings, producing artifacts when resolution changes. CTO instead learns in continuous function space, letting a single model handle multiple sampling rates—across organs and diagnostic protocols—without retraining.
CTO introduces two architectural innovations: a dual-domain NO that processes both sinogram and image spaces for complementary spatial-frequency info, and rotation-equivariant DIScrete-COntinuous (DISCO) convolutions that exploit CT's rotational structure. In multi-resolution benchmarks, CTO beats CNNs by over 3.4dB PSNR and outperforms state-of-the-art diffusion models with 500x faster inference and a 3dB gain. It also maintains strong performance under cross-dataset transfer and noisy sinogram conditions. The code is open-sourced, marking neural operators as a practical paradigm for flexible, discretization-agnostic CT reconstruction.
- CTO is the first neural operator framework for CT reconstruction, generalizing across acquisition sampling rates without retraining
- Outperforms CNNs by >3.4dB PSNR in multi-resolution settings and beats diffusion models with 500x faster inference plus 3dB average gain
- Dual-domain NO architecture and rotation-equivariant DISCO convolutions enable robust cross-dataset transfer and noisy sinogram handling
Why It Matters
One adaptable CT model could cut radiation doses and scan times while working across clinical protocols and scanners, enabling faster, cheaper imaging.