Programa del congreso

Noise in X-ray imaging affects both image quality and diagnostic accuracy. Deep learning-based denoising methods have shown potential for dose reduction, but their performance depends on the characteristics of the training data. Synthetic low-dose images are commonly used within self-supervised training strategies, yet standard noise models often use oversimplified models of the imaging chain, that fail to capture the spatial and frequency properties of quantum and electronic noise in flat-panel detectors.This work illustrates the impact of noise model accuracy on the performance of deep denoising methods trained in a self-supervised fashion. A novel model-based noise synthesis framework capable of generating dose-dependent, high-fidelity noise for any input image is presented and its performance for training of denoising operators is compared against conventional models of Poisson noise. The model leverages of a comprehensive model of the frequency and spatial propagation of quantum and electronic noise along the imaging chain. The synthetic noise model was validated against experimental measurements across clinically relevant protocols. The effect of noise fidelity on network training was evaluated using a U-Net CNN trained with images corrupted either by the proposed noise model or by a simpler, uncorrelated Poisson noise model. Networks trained with high-fidelity noise achieved higher SSIM and PSNR and better preservation of structural details attributable to better mitigation of correlated mid-frequency noise, highlighting the need of highly realistic synthetic models for training of deep denoising operators aimed to be deployed in clinical scenarios.