Feasibility of Deep Learning-Guided Attenuation and Scatter Correction of Whole-Body 68Ga-PSMA PET Studies in the Image Domain

Objective This study evaluates the feasibility of direct scatter and attenuation correction whole-body 68Ga-PSMA PET images in image domain using deep learning. Methods Whole-body 399 subjects were used to train a residual learning model, taking non–attenuation-corrected (PET-nonAC) as input CT-based attenuation-corrected (PET-CTAC) target (reference). Forty-six an independent validation dataset. For validation, synthetic learning–based assessed considering corresponding PET-CTAC reference. The evaluation metrics included mean absolute error (MAE) SUV, peak signal-to-noise ratio, structural similarity index (SSIM) whole body, well different regions namely, head neck, chest, abdomen pelvis. Results learning–guided produced comparable visual quality images. It achieved MAE, relative (RE%), SSIM, ratio 0.91 ± 0.29 (SUV), −2.46% 10.10%, 0.973 0.034, 48.171 2.964, respectively, within external largest RE% was observed neck region (−5.62% 11.73%), although this exhibited highest value SSIM metric (0.982 0.024). MAE (SUV) body less than 2.0% 6%, indicating acceptable performance model. Conclusions work demonstrated with clinically tolerable errors. technique has potential performing on stand-alone or PET/MRI systems.