Background: Positron emission tomography (PET) is a key tool for quantitative brain imaging, but its image quality and quantitative reliability are strongly dependent on injected radiotracer activity and acquisition time. Reducing injected dose or acquisition time lowers radiation exposure but increases image noise, while partial volume effects (PVE) further degrade signal accuracy, especially in cortical regions critical for neurodegenerative disease assessment. Conventional partial volume correction (PVC) methods such as Iterative Yang can improve quantification but are highly sensitive to noise and perform poorly in low count settings, while also relying on accurate anatomical information, typically derived from MRI, which may be unavailable, misregistered, or impractical in routine clinical workflows; in parallel, most deep learning (DL) approaches reported to date focus mainly on denoising and do not jointly address PVE. A method capable of simultaneously mitigating low count noise and PVE is needed. This study aimed to evaluate whether a DL approach can generate full count, PVC brain PET images from fast or low count non PVC scans, using both clinical and phantom datasets.

Results: Using clinical and phantom brain PET datasets with ¹⁸ F-FDG and ¹⁸ F-Florbetapir, we compared low count PET images at 10% of full counts with images processed by Iterative Yang PVC and a Swin Unetr based DL-PVC model. The DL-PVC method consistently outperformed low count non-PVC PET across all evaluations. It achieved higher structural similarity index (SSIM) by about 2%, higher peak signal to noise ratio (PSNR) by about 11%, and lower relative root mean square error (rRMSE) by about 4%, indicating improved image quality and quantitative accuracy. In the clinical test sets, DL-PVC further showed clear gains for both tracers, with SSIM improvements of 2.1% for ¹⁸ F-FDG and 2% for ¹⁸ F-Florbetapir, PSNR increases of 21.6% and 22.3% respectively, and substantial rRMSE reductions of 50% for ¹⁸ F-FDG and 44% for ¹⁸ F-Florbetapir.

Conclusions: Our study introduced a DL based PVC method based on Transformer and Unet architectures. By combining phantom and clinical data through a transfer learning strategy, the network benefits from the phantom ground truth activity maps while also learning the complexities present in real clinical PET. With this approach, it improves PET image quality and quantitative agreement with a high count reference, and at inference it requires only PET input, so it can be deployed without MR images for PVC.