Noise Suppression in Ultra-Low-Dose CT for LAFOV PET/CT via Unpaired CycleGAN Training

Résumé fourni par la source

Reducing radiation dose in positron emission tomography / computed tomography (PET/CT) is critical, and recent advances in ultra-low-dose (ULD)-CT protocols have enabled up to a 98% reduction in CT radiation exposure. However, this comes at the cost of increased image noise and reduced anatomical detail, which may hinder visual interpretation in some clinical indications. While low-dose (LD) CT, typically used in PET/CT, is not of diagnostic quality, it provides adequate anatomical information for attenuation correction and localization. This study introduces a deep learning-based denoising approach using a cycle-consistent generative adversarial network (cGAN) to enhance ULD-CT image quality. A total of 70 patient scans were used: 51 for training and internal validation and 19 for external validation. Due to slight anatomical mismatches, the model was trained on unpaired two-dimensional slices. Performance was assessed using signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), and multiple image similarity metrics. Results showed that synthetic cGAN-CT images improved SNR and CNR across the lung, liver, and bone, with relative differences to LD-CT under 16%, while ULD-CT deviated by over 60%. Image similarity metrics confirmed that cGAN-CT achieved structural quality comparable to LD-CT. This approach supports substantial dose reduction while preserving image quality for clinical PET/CT interpretation.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.