Optimization of PET Image Reconstruction for Enhanced Image Quality in Various Tasks Using a Conventional PET Scanner

Positron emission tomography (PET) imaging requires high‐quality yet rapid reconstruction to ensure clinical effectiveness, as these reconstructions enable timely and accurate diagnosis, guide treatment decisions, and reduce the risk of delayed interventions in critical clinical settings. This study introduces a deep learning‐based method that employs conditional generative adversarial networks (cGANs) for direct sinogram‐to‐image PET reconstruction. A dual approach was used: simulation experiments with Zubal phantoms, which provided a controlled and reproducible environment to test the reconstruction accuracy and robustness, and validation with real patient datasets, ensuring the method’s applicability and effectiveness in clinical settings. The primary objective was to evaluate the ability of the cGAN‐based method to enhance image quality, reduce noise, and improve reconstruction speed compared to conventional algorithms, such as maximum likelihood expectation maximization (MLEM) and total variation (TV). The methodology involved training a U‐net‐based generator and a whole‐image discriminator iteratively to reconstruct PET images with superior resolution and accuracy. Key outcome measures included bias, variance, structural similarity index (SSIM), and relative root mean square error (rRMSE), as these metrics effectively quantify image fidelity, noise levels, and structural accuracy, which are critical for evaluating the clinical reliability and precision of reconstructed PET images. The results showed that the proposed method achieved significant improvements in image clarity, noise suppression, and computational efficiency, outperforming the traditional techniques. These findings highlight the potential of cGAN‐based reconstruction in improving diagnostic accuracy and clinical workflow.