Relevancy aware cascaded generative adversarial network for LSO-transmission image denoising in CT-less PET

Purpose. Achieving high-quality PET imaging while minimizing scan time and patient radiation dose presents significant challenges, particularly in the absence of CT-based attenuation maps. Joint reconstruction algorithms, such as MLAA and MLACF, partially address these challenges but often result in noisy and less reliable images. Denoising these images is critical for enhancing diagnostic accuracy.Approach. This study introduces a novel cascaded relevancy-aware Generative Adversarial Network (reGAN) to improve the denoising and diagnostic reliability ofμ-maps derived from joint reconstruction algorithms, ultimately aimed at enhancing PET imaging quality. The reGAN architecture employs a cascaded design incorporating UPlus GAN modules, relevancy mapping, and contextual attention mechanisms. The model was trained using PET/CT data from 16 patients, with MLAA and MLACF-derivedμ-maps as input and CT-basedμ-maps as the ground truth. Performance was evaluated using metrics such as SSIM, PSNR, VIF, and MSE. Comparative studies were conducted against other popular 2D and 3D GAN architectures.Results. The proposed reGAN achieved the highest SSIM (0.91 for MLAA and 0.93 for MLACF), PSNR (34.7 dB for MLAA and 36.2 dB for MLACF), and VIF (0.89 for MLAA and 0.91 for MLACF), while maintaining the lowest MSE (0.021 for MLAA and 0.018 for MLACF). Qualitative analysis demonstrated that reGAN preserved fine details, particularly in bony structures, and reduced artifacts effectively. Additionally, relevancy maps provided pixel-wise confidence indicators, further aiding interpretability and diagnostic reliability.Conclusion. reGAN presents a robust approach to medical image denoising, combining advanced generative modeling with diagnostic confidence metrics. The proposed method constitutes a viable approach for achieving quantitative accuracy in low-dose PET imaging in the absence of CT.