Deep learning-based approaches for attenuation correction in [18F]FDG PET: Current advances and future directions.

BACKGROUND: Accurate attenuation correction (AC) is essential for quantification in [18F]FDG PET imaging. However, CT or MRI-based methods may introduce artifacts, prolong scan duration, or increase radiation exposure. Deep learning-based AC (DLAC) has emerged as an alternative, enabling the generation of attenuation-corrected PET images. METHODS: This review examines advancements in DL driven AC from January 2019 to August 2025, focusing on whole-body, brain, and chest PET applications, with emphasis on quantitative performance and architectural developments. RESULTS: Early convolutional neural networks (CNNs) typically achieved SUV biases of 5-7%, though recovery of structural details remained limited. U-Net architectures enhanced DL model performance, yielding SSIM values of 0.97-0.99, PSNR of 35-36 dB, and SUV biases below 5%. GAN-based architectures improved lesion quantification, producing PSNR of 35-37 dB, SSIM of approximately 0.98, and SUV errors of 3-6%. Vision Transformers (ViT), with their advanced structure, attained PSNR near 38 dB, SSIM close to 0.99, and SUV biases under 5%, particularly in chest and whole-body datasets. More recently, diffusion-based models have exhibited excellent performance, with PSNR of 36-38 dB, SSIM around 0.98, and SUV biases consistently below 3%, alongside enhanced robustness to noise, motion, and low-count acquisitions. CONCLUSION: DLAC provides clinical benefits, including reduced radiation exposure, shorter acquisition times, artefact reduction, and reliable quantitative accuracy. Despite persistent challenges related to dataset heterogeneity, generalizability, and clinical validation, DLAC promises to enable accurate, low-dose, and fully CT/MRI free PET imaging in future clinical practice.