Quantitative evaluation of image quality restoration for short-acquisition brain PET using deep learning models and a...

This study evaluated the performance of U-Net and Nested U-Net in enhancing brain positron emission tomography (PET) image quality from 1-min acquisitions using a dedicated copper-based three-dimensional-printed striatum phantom. The phantom was fabricated with copper filament, filled with Fluorine-18, and imaged using a clinical digital PET/computed tomography scanner. A total of 500 paired datasets, consisting of 1-min input and 10-min reference images, were divided into training, validation, and test sets in an 8:1:1 ratio. The models were trained in PyTorch using the Adam optimizer. Performance was evaluated using the computational time, root mean square error (RMSE), peak signal-to-noise ratio (PSNR), universal quality index (UQI), learned perceptual image patch similarity (LPIPS), contrast-to-noise ratio (CNR), and coefficient of variation (COV). The U-Net (15.9 h) required approximately 2.3 times less training time than the Nested U-Net (37.2 h). Both models substantially improved RMSE, PSNR, UQI, LPIPS, and CNR compared with the 1-min input images (p < 0.001). The U-Net achieved slightly superior RMSE, PSNR, and UQI compared to the Nested U-Net (p < 0.01), whereas both showed comparable LPIPS, CNR, and COV values. These results demonstrate that U-Net provides an effective balance between computational efficiency and quantitative accuracy for phantom-validated image quality restoration under short-acquisition PET conditions, which is directly relevant to applied PET imaging.