Researchers say CTO reconstructs CT images across sampling rates without retraining

A research team introduces Computed Tomography neural Operator (CTO) for sparse-view CT, which uses fewer X-ray projections. Taking fewer projections cuts radiation dose and scan time, the team says, but makes image reconstruction harder.

The researchers say existing deep learning approaches need a separate model for each sampling rate, while CTO handles any sampling rate without retraining. They report average PSNR gains of 3.42 dB over a traditional unrolled CNN variational network and 6.02 dB over diffusion models, alongside 500-fold faster performance than diffusion models. PSNR is an image-quality metric.

The announcement links to the paper, project page and code.