Abstract
The novel deep learning image reconstruction (DLIR) is known to change its image quality characteristics according to object contrast and image noise. In clinical practice, computed tomography (CT) image noise is usually controlled by tube current modulation (TCM) to accommodate changes in object size. This study aimed to evaluate the image quality characteristics of DLIR for different object sizes when in-plane noise is controlled by TCM. We used Mercury 4.0 phantoms with different object sizes. Phantom image acquisition was performed on a GE Revolution CT system to investigate the impact of the DLIR algorithm compared to standard reconstructions: filtered back projection (FBP) and hybrid iterative reconstruction (hybrid-IR). For image quality evaluation, the noise power spectrum (NPS), task-based transfer function (TTF), and detectability index (d') were determined. The NPS of DLIR was very similar to that of FBP, and the information in the high-frequency region was maintained. In terms of TTF, DLIR showed higher resolution than hybrid-IR at low- to medium-contrast (Δ50, Δ90HU), but not necessarily higher than FBP. At the simulated contrast and lesion size, DLIR showed higher detectability than hybrid-IR, regardless of the phantom size. In this study, we evaluated a novel DLIR algorithm by reproducing clinical behaviors. The findings indicate that DLIR produces higher image quality than hybrid-IR regardless of the phantom size, although it depends on the reconstruction strength.
Machine Friendly Machine Learning: Interpretation of Computed Tomography Without Image Reconstruction