Optimizing acquisition time and injected dose in ¹⁸F-FDG PET/CT imaging using deep learning: enhancing image protocol...

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Artificial Intelligence

Background:

• Optimizing 18F-FDG PET/CT protocols is crucial for balancing image quality, radiopharmaceutical use, and patient radiation exposure.
• Shortened acquisition times in PET/CT can compromise image quality and diagnostic accuracy.

Purpose of the Study:

• To evaluate the effectiveness of deep learning algorithms in reconstructing high-quality 18F-FDG PET/CT images from shortened acquisition times.
• To assess the impact of deep learning on image quality, lesion detectability, and radiopharmaceutical optimization.

Main Methods:

• A residual U-Net architecture with data augmentation was used to reconstruct 18F-FDG PET/CT images from shortened acquisition times (30-90s) and half-dose conditions in 322 patients.
• Image quality was assessed using Likert scales, inter-reader consistency (Cohen's kappa), and quantitative metrics (PSNR, SSIM, MAE, MSE, RMSE).
• Lesion detectability was evaluated by senior nuclear medicine physicians, and statistical analyses included Wilcoxon signed-rank tests and kappa statistics.

Main Results:

• Deep learning significantly improved image quality scores (Likert scale) and kappa values in half-dose and shortened acquisition scenarios (p < 0.05).
• Quantitative metrics showed marked improvements, with SSIM increasing from 0.75 to 0.87 and PSNR from 37.65 dB to 41.37 dB in half-dose conditions.
• Lesion detectability, particularly for small lesions, improved by up to 16.3% under shortened acquisition conditions, enhancing the detection of pathological features.

Conclusions:

• Deep learning models can effectively reconstruct high-quality 18F-FDG PET/CT images from shortened acquisition times and reduced doses.
• AI-enhanced imaging protocols show potential for improving scan efficiency, reducing radiopharmaceutical usage, and minimizing patient radiation exposure without compromising diagnostic information.