Abstract

The reliability of deep learning (DL) denoising models in low-dose computed tomography (CT) is vital to prevent hallucinated structures and ensure clinical trust. This study introduces AntiHalluciNet, a novel framework designed to detect spurious structural components in residual noise from DL denoising models. Trained on paired structure-embedded and pure noise images, AntiHalluciNet identifies unwanted artifacts using the Residual Structure Index (RSI), a new metric for quantifying prediction confidence. Its ability to assess denoised image fidelity using only noise residuals offers a practical alternative to the Structural Similarity Index (SSIM), which requires reference images. Applied to audit RED-CNN, CTformer, and ClariCT.AI (version 1.2.3), AntiHalluciNet achieved RSI values of 0.57 ± 0.31 for structure-embedded noise and 0.02 ± 0.02 for pure noise in a 50% low-dose dataset (p < 0.0001). For degraded images, RSI outperformed SSIM across 25%–100% degradation levels, showing greater differentiation. Distinct RSI distributions 0.28 ± 0.06 (RED-CNN), 0.21 ± 0.06 (CTformer), and 0.15 ± 0.03 (ClariCT.AI) highlight AntiHalluciNet’s precision in auditing DL denoising models, enhancing clinical reliability by detecting structural artifacts in residual noise.