Sharp but Wrong? Fidelity and Hallucination in Deep CT Super-Resolution with Analytic Phantoms

Abstract Deep super-resolution (SR) techniques could potentially recover the resolution that X-ray computed tomography (CT) sacrifices to reduce radiation dose and scanning time. However, in CT imaging, generating details that appear plausible but are not physically present represents a critical failure that standard evaluation metrics cannot detect.In this study, we train two deep-learning models to upsample CT slices: a pixel-loss residual network (EDSR) and an adversarially trained network (ESRGAN). Both models are evaluated in terms of reconstruction fidelity and their tendency to hallucinate image details.To assess hallucinated details, we use analytical phantoms containing line-pair gratings and pores with precisely known frequencies and diameters. The evaluation includes held-out frequency bands, a documented blur–decimation–noise model, and a validated spectral measurement method that classifies each grating as resolved, washed out, aliased, or hallucinated.At ×4 upscaling, all neural networks improve PSNR by more than 6 dB compared with bicubic and Lanczos interpolation. ESRGAN achieves the best perceptual score (LPIPS = 0.015 compared with 0.046 for EDSR), at the cost of approximately 2 dB in reconstruction fidelity. However, ESRGAN hallucinates 61% of the gratings that cannot be represented in the input image, whereas interpolation methods produce no hallucinated gratings.An adversarial-weight sweep demonstrates that perceptual loss alone is responsible for these invented details. Pixel-loss networks also exhibit limitations, replacing held-out gratings just below the Nyquist limit with their aliased counterparts learned during training.At ×2 upscaling, hallucination rates decrease substantially but do not disappear entirely, with ESRGAN still hallucinating 27% of the gratings. Interestingly, pores and thin lines that were never encountered during training are successfully recovered by all networks.All experimental data are synthetic and reproducible using a specified random seed.