Abstract: With the increasing popularity and clinical adoption of deep learning in CT image reconstruction, two distinct approaches have emerged under the concept of 'deep' reconstruction: Deep Learning Reconstruction (DLR) and Deep Iterative Reconstruction (DIR). Despite falling under the umbrella of 'deep' reconstruction, DLR and DIR differ in technical principle, clinical applicability, and reconstruction performance. This review aims to provide a clinically oriented overview of these two methods, emphasizing their coexistence and differentiated roles on a 320-row CT scanner platform, offering radiologists insights for clinical practice as well as inspirations for future research. On this platform, DLR and DIR represent complementary strategies in clinical practice, where DLR is implemented as a cardiac-specific algorithm and DIR for other bodyparts. By summarizing representative clinical applications, we highlight the advantages of DLR in cardiac CT and strengths of DIR across chest, abdominal, vascular, and perfusion CT imaging. Quantitative evidence from recent studies demonstrates consistent improvements of both DIR and cardiac-specific DLR over routine Hybrid Iterative Reconstruction (HIR). Their complementary characteristics also suggest potential benefits when applied in multi-region CT imaging. In addition, the clinically valuable image features of DIR that merit further investigation, as well as other technical considerations relevant to 'deep' reconstructions are discussed.
Zhou et al. (2026) studied this question.