CT scout scans with deep learning for ultra-low dose attenuation correction in PET

Most clinical PET scanners integrate a CT for anatomical correlation and PET data corrections. However, when there is a spatial mismatch between PET and CT due to respiratory or gross body motion, CT-based attenuation correction (AC) results in image artefacts and quantification biases. Deep learning (DL) shows promise for CT-less AC, particularly when trained on specific organ regions. We propose to use non-corrected (NC) PET and scout images (= dual-view localizer radiographs acquired at the start of PET/CT scans) as inputs to a dual-branch network with the goal to extract complementary, modality-specific features. While scouts are routinely used for scan planning, the 2D projection images contain anatomical tissue density information that can be leveraged by a neural network for AC in PET, potentially precluding the need for the full CT scan. Our proposed approach aims to optimize a DL framework for a variety of tracer studies in long axial field-of-view PET which currently presents a challenging case for DL-based AC methods. Evaluations on 18 F