Evaluation of deep-learning iterative reconstruction combined with high-frequency kernels in CT: a task-based image quality study.

objectiveTo evaluate the image quality of a novel deep-learning iterative reconstruction (DLIR) algorithm combined with high-frequency kernels compared with iterative reconstruction (IR) algorithms through the assessment of the detectability of small high-contrast lesions. MATERIALS AND

methodsThree image-quality phantoms were scanned with doses from 0.8 to 15 mGy. Images were reconstructed using ASIR-V0, ASIR-V80, and DLIR (Low, Medium, High), with an edge-enhancing kernel for a chest protocol and a sharp kernel for a spine protocol. A dedicated cubic phantom was used to assess axial, longitudinal resolution, and noise power spectra. Contrast of calcium-based lesions of 3 and 5 mm in diameter was assessed at three concentrations (200, 400, and 800 mg/cc) using two anthropomorphic chest and abdominal phantoms. Detectability was evaluated using a non-prewhitening with eye filter model observer.

resultsIn-plane spatial resolution was stable across DLIR strength levels and slightly better than ASIR-V0, while longitudinal resolution did not depend on algorithms. DLIR markedly reduced image noise, below ASIR-V80 for DLIR-High with both kernels. While IR preserved the contrast of lesions, it decreased with DLIR strength for the small lesion 200 mg/cc-3mm. The highest detectability was achieved with DLIR-High, except for the HA200-3mm below 3 mGy for the chest protocol and 7 mGy for the spine protocol.

conclusionDLIR combined with high-frequency kernels reduced image noise while preserving spatial resolution and FBP-like noise texture, outperforming IR in detectability for most lesion sizes, concentrations, and dose levels. DLIR reduced contrast for small and low-concentration lesions, especially at low doses, which partially counterbalanced its detectability advantage over IR. KEY POINTS: Question This study evaluates whether deep-learning reconstruction with high-frequency kernels improves the detectability of small high-contrast lesions compared to iterative reconstruction. Findings The new deep-learning reconstruction reduces image noise while preserving fine anatomical detail and improving in-plane spatial resolution. RELEVANCE STATEMENT: The new deep-learning reconstruction combined with high-frequency kernels improves detection of small lesions in comparison with iterative reconstruction at standard doses, but very small or low-density lesions may appear less visible at lower dose levels.