Optimizing pediatric chest CT: superior image quality and lower radiation dose with deep learning reconstruction

Objectives To compare image quality and radiation dose between deep learning reconstruction (DLIR) and hybrid iterative reconstruction (HIR) algorithms in unenhanced pediatric chest CT. Materials and methods This hybrid prospective-retrospective study included 142 pediatric patients (<16 years) who underwent single-phase unenhanced chest CT between 2021 and 2024. The DLIR cohort was prospectively enrolled using a new-generation scanner, whereas an age- and indication-matched HIR cohort was retrospectively identified. Subjective image quality was evaluated by two pediatric radiologists using a 5-point Likert scale, with inter-reader agreement assessed by weighted Cohen’s kappa. Objective image quality was quantified by attenuation and noise measurements across thoracic structures. Comparisons were also stratified by age groups (0–5, 6–10, and 11–16 years). Radiation exposure was assessed using CTDIvol, DLP, and estimated effective dose (ED), and statistical analysis was applied to compare reconstruction techniques. Results Subjective image quality was comparable between DLIR and HIR, with substantial inter-reader agreement (κ = 0.78). Although DLIR did not consistently outperform HIR in visual assessment, quantitative analysis demonstrated significantly lower image noise with DLIR in the lung parenchyma (63 vs 81 HU, p = 0.017), trachea (50 vs 66 HU, p < 0.001), and paraspinal muscle (12 vs 15 HU, p < 0.001), while attenuation values were comparable across techniques. Radiation exposure was markedly reduced with DLIR, showing lower CTDIvol (1.6 vs 4.3 mGy), DLP (43 vs 145 mGy·cm), and median effective dose (0.70 vs 2.51 mSv, p < 0.001) compared with HIR. Conclusions Deep learning reconstruction appears to enable a substantial reduction in radiation exposure while maintaining or improving diagnostic image quality in pediatric chest CT.