Deep learning for CT-less attenuation correction synthesis in non-small cell lung cancer: A proof-of-concept...

To develop and rigorously validate a deep learning framework for CT-free positron emission tomography (PET) attenuation correction in non-small cell lung cancer (NSCLC), with a comprehensive assessment of both quantitative and physical integrity to establish a proof-of-concept for its technical feasibility. A U-Net was trained on a multi-center dataset of 100 NSCLC patients to synthesize attenuation-corrected (AC) from nonattenuation-corrected (NAC) PET data. The framework's performance was evaluated on an independent 30-patient test set. The multi-faceted validation protocol included global image fidelity (peak signal-to-noise ratio [PSNR], structural similarity index measure [SSIM]), quantitative lesion integrity (SUV_max/SUV_mean bias, dice similarity coefficient [DSC]), and fundamental physical characteristics (contrast recovery, spatial resolution). Bland-Altman analysis and paired statistical tests were used. The framework generated images with state-of-the-art global fidelity (PSNR: 36.2 dB; SSIM: 0.967). For lesion-specific analysis, the model introduced a negligible and statistically insignificant mean bias in both SUV_max (+0.72%;p=0.067). While SUV_mean showed a statistically significant difference due to per-patient aggregation, the absolute bias remained clinically negligible (-1.23%;p=0.005). Spatial accuracy was excellent (mean DSC: 0.89). Critically, physical performance was preserved, with 99.2% contrast recovery and no significant degradation in spatial resolution. Deep learning-based attenuation correction synthesis is a highly promising proof-of-concept for CT-less thoracic PET. By demonstrating quantitative fidelity and preserved image-derived physical characteristics, this work establishes a foundational step toward a prospective emission-only workflow, which has the future potential to enhance patient safety by eliminating CT-associated radiation dose.