Image Quality Advancements in Low-Dose Pediatric CT Using Super-Resolution Deep-Learning Reconstruction

The objective of the study is to assess the impact of super-resolution deep-learning reconstruction (SR-DLR) on image quality in low-dose thin-slice pediatric abdominal CT. Thirty-eight children (< 10 years; median age, 2.0 [IQR: 0.0-3.8] years) who had available low-radiation abdominal CT data were retrospectively analyzed. The mean size-specific dose estimate was 1.71 ± 0.38 mGy. From the raw datasets, 0.5-mm images were generated using SR-DLR developed for helical body imaging as well as hybrid-iterative reconstruction (hybrid-IR) and conventional deep-learning reconstruction (C-DLR), with matrices of 1024, 512, and 512, respectively. Quantitative assessments included measurements of image noise and contrast-to-noise ratio (CNR). Image noise characteristics were evaluated using the noise power spectrum (NPS), while edge sharpness was evaluated based on the edge-rise slope (ERS). For qualitative evaluation, noise magnitude, texture, sharpness, and small structure delineation were ranked among three reconstructions (1 = worst, 3 = best). Diagnostic confidence was scored with a five-point scale (1 = undiagnostic, 5 = most confident). Quantitative and qualitative data were compared among hybrid-IR, C-DLR, and SR-DLR. SR-DLR showed the greatest noise reduction and the highest CNR, outperforming the other reconstructions (p < 0.001). SR-DLR also achieved the highest ERS and maintained the lowest noise throughout the full range of spatial frequencies without a shift toward lower frequency. Subjective scores aligned with these quantitative findings, with SR-DLR consistently achieving the highest scores across all qualitative metrics (all p < 0.001). In conclusion, SR-DLR with 1024-matrix for helical body imaging enhanced sharpness and reduced noise in low-dose thin-slice pediatric abdominal CT, outperforming hybrid-IR and C-DLR in small structure delineation and overall diagnostic quality.