Clinical consequences of deep learning image reconstruction at CT

Deep learning reconstruction (DLR) offers a variety of advantages over the current standard iterative reconstruction techniques, including decreased image noise without changes in noise texture and less susceptibility to spatial resolution limitations at low dose. These advances may allow for more aggressive dose reduction in computed tomography imaging while maintaining image quality and diagnostic accuracy. However, the performance of DLRs is impacted by the type of framework and training data used. In addition, the patient size and clinical task being performed may impact the amount of dose reduction that can be reasonably employed. Multiple DLRs are currently Food and Drug Administration (FDA) approved with a growing body of literature evaluating performance throughout this body; however, continued work is warranted to evaluate a variety of clinical scenarios to fully explore the evolving potential of DLR. Depending on the type and strength of DLR applied, blurring and occasionally other artefacts may be introduced. DLRs also show promise in artefact reduction, particularly metal artefact reduction. This commentary focuses primarily on current DLR data for abdominal applications, current challenges, and future areas of potential exploration.