AI Denoising Cuts PET Scan Time but Struggles with Radiomic Features

Deep learning models can sharply cut the noise in respiratory-gated PET/CT images acquired at a fraction of the usual scan time, but they come with a trade-off that clinicians may not have anticipated: the artificial-intelligence-enhanced images are less faithful to the original data than simply using the shorter scans themselves when it comes to radiomic features. That is the central finding of a new clinical evaluation published in EJNMMI Physics.

The study enrolled 90 patients who underwent respiratory-gated whole-body 18F-FDG PET/CT scans, with the respiratory-gating trigger phase width set to 35%. Using list-mode data from full-acquisition-time PET scans of 240 seconds per bed position, the researchers reconstructed four sets of low-acquisition-time PET images at 30, 60, 90, and 120 seconds. They then trained two types of deep learning models, a U-net and a Generative Adversarial Network (GAN), to enhance each of the shortened datasets.

The evaluation went beyond simple image aesthetics. The team assessed denoising performance using clinical measures including subjective Likert scores for image quality, metabolic metrics, lesion detection, and the stability of radiomic features, the quantitative texture descriptors increasingly used in oncology research.

Both the GAN and U-net models improved subjective Likert scores for image quality and lesion conspicuity while maintaining consistency in metabolic metrics. For the 120-second enhanced images, quantitative agreement with the full-acquisition scans was strong: the mean absolute percentage error for lesion SUVpeak was 5.6% for GAN-enhanced images, with a bias of −0.19 ± 0.20, and 5.5% for U-net-enhanced images, with a bias of −0.09 ± 0.27.

Lesion detection in the enhanced images rose with acquisition time, but the two architectures behaved differently. GANs consistently improved lesion detection in low-acquisition-time PET images, whereas U-nets offered no benefit at 60 seconds or longer. Compared with 120-second free-breathing PET, GAN enhancement delivered superior overall lesion detection at both 90 seconds (p = 0.035) and 120 seconds (p < 0.001). The 120-second low-acquisition-time images also outperformed free-breathing PET (p < 0.001), while 90-second low-acquisition-time images and 120-second U-net-enhanced images showed comparable performance (p = 0.749 and p = 0.086, respectively).

A striking safety-relevant difference emerged between the two model types. Nine U-net-enhanced images showed false-positive liver uptake, a spurious signal that could complicate clinical interpretation, while the GAN-enhanced images remained free of false positives. This suggests that generative models may be more reliable when the priority is avoiding artifacts that mimic disease.

However, the picture was less favorable for radiomics. Most radiomic features from the low-acquisition-time PET images showed greater consistency with the full-acquisition-time features than the features extracted from AI-enhanced images did. In other words, the deep learning enhancement, while visually appealing and quantitatively consistent for standard metabolic metrics, did not reliably restore the radiomic signature of the original scans.

The authors conclude that AI image enhancement, particularly generative models, can effectively reduce noise in low-acquisition-time PET images, improve lesion detection, and maintain metabolic metrics, but that these models demonstrate limitations in restoring radiomic features. The findings validate the clinical benefits of AI-based PET image enhancement while acknowledging their constraints for radiomics analysis, a caveat that matters for any workflow intending to feed enhanced images into radiomic feature extraction pipelines.

For nuclear medicine departments weighing faster scans against diagnostic confidence, the study offers a nuanced message: shorter acquisition times paired with GAN enhancement can approach the lesion detection performance of longer scans, and even outperform free-breathing acquisitions, but the technology is not a universal substitute for the full-acquisition gold standard, particularly where radiomic analysis is concerned.

Subject of Research: Deep learning enhancement of respiratory-gated 18F-FDG PET/CT images acquired at reduced acquisition times

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

Article References: Wang, G., Shen, J., Liu, Y., Wang, C., Wang, H., Liu, K., Yang, S., Yu, F., Su, X., & Luo, W. (2026). Clinical evaluation of deep learning-enhanced respiratory gated 18F-FDG PET/CT images. EJNMMI Physics. https://doi.org/10.1186/s40658-026-00947-4

Image Credits: AI Generated

DOI: 10.1186/s40658-026-00947-4

Keywords: PET/CT, deep learning, GAN, U-net, radiomics, lesion detection, image denoising, respiratory gating, 18F-FDG, SUV, nuclear medicine, image quality