Clinical evaluation of deep learning-enhanced respiratory gated 18F-FDG PET/CT images

Abstract Objectives This study aims to comprehensively assess the value of deep learning (DL) models in improving the image quality of respiratory-gated PET/CT and explore the impact of PET acquisition time on the quality of model-predicted images. Methods We enrolled 90 patients who underwent respiratory-gated PET/CT whole-body scans, and the respiratory-gating trigger phase width was set to 35%. Using list-mode data from full-acquisition-time PET (FAT-PET, 240 s/bed), we reconstructed four sets of low-acquisition-time PET (LAT-PET) images at 30, 60, 90, and 120 s. We trained U-net and Generative Adversarial Network (GAN) models to enhance each dataset. The evaluation of denoising performance considered clinical aspects including subjective scores, metabolic metrics, lesion detection, and radiomic feature stability. Results Both GAN and U-net models improved subjective Likert scores for image quality and lesion conspicuity while maintaining metabolic metric consistency. 120 s GAN and U-net enhanced images showed strong quantitative consistency with FAT-PET, with mean absolute percentage error (MAPE) of SUV peak for lesions being 5.6% (bias, − 0.19 ± 0.20) and 5.5% (bias, − 0.09 ± 0.27), respectively. The detection of lesions in GAN- and U-net-enhanced images was positively correlated with acquisition time. GANs consistently enhanced lesion detection in LAT-PET images, whereas U-nets provided no benefit for 60 s or higher acquisition times. Compared with 120 s free-breathing PET (FB-PET), GAN-enhancement provided superior overall lesion detection at 90 s ( p = 0.035) and 120 s ( p < 0.001); 120 s LAT-PET also outperformed FB-PET ( p < 0.001), while 90 s LAT-PET and 120 s U-net-enhanced images showed comparable performance ( p = 0.749 and p = 0.086, respectively). We also found nine U-net enhanced images showed false-positive liver uptake, while GAN-enhanced images remained free of false positives. Additionally, most radiomic features from LAT-PET images showed greater consistency with FAT-PET features compared to those from AI-enhanced images. Conclusion AI image enhancement—particularly generative models—can effectively reduce noise in LAT-PET images, improve lesion detection, and maintain metabolic metrics. However, they demonstrate limitations in restoring radiomic features. These findings validate the clinical benefits of AI-based PET image enhancement while acknowledging their constraints in radiomics analysis.