Purpose or Case Report: Patients with congenital cardiac disease are a vulnerable population who require early and repeat CT imaging. However, decreasing ionizing radiation dose in pediatric CT increases image noise. We evaluated a self-supervised deep learning denoising model integrating sparse coding with a modified Vision Transformer (SC-mViT) compared with a non-local means (NLM) algorithm using quantitative and qualitative metrics.
Methods & Materials: 29 pediatric congenital cardiac CT scans were denoised with SC-mViT model and NLM algorithm. Quantitative image quality metrics including average global noise index (AGNI) and average NPS frequency (NPSfav) were assessed across the 29 scans: without denoising, using SC-mViT model, and using NLM. Quantitative image quality metrics between original and denoising conditions were compared using paired t-tests.

The three image sets were randomized and reviewed by 4 blinded pediatric radiologists. Readers evaluated image quality and noise on a 1–10 Likert scale. A coronary artery detection task with confidence ratings (none, low, moderate, high) was also asked. The intraclass correlation coefficient was calculated for the Likert responses. Agreement metrics were assessed using Krippendorff’s alpha.
Results: Both SC-mViT and NLM significantly reduced AGNI versus original. NLM achieved the lowest AGNI, 4.54 HU lower than SC-mViT on average (95% CI: −5.09 to −4.00; p < 0.001). Both denoising methods reduced the average NPS frequency with SC-mViT showing higher NPSfav than NLM.

NLM image quality and noise Likert scores showed the poorest agreement amongst 4 readers (ICC 0.370 and 0.216 respectively). SC-mViT showed highest reader agreement (ICC 0.618, 0.546). SC-mViT image noise was rated significantly better than original by all readers and significantly better than NLM by 2 readers. SC-mViT image quality was rated significantly better than original by 2 readers and significantly better than NLM by 2 readers. Inter-reader agreement for artery detection confidence was fair, slightly higher for SC-mViT and original than NLM (Kripendorff’s alpha 0.802 vs 0.635). No significant differences were observed for artery detection confidence.
Conclusions: SC-mViT effectively reduces image noise while preserving diagnostic quality in low-dose pediatric cardiac CT. SC-mViT produces a more favorable noise texture and was preferred by radiologists. Our study supports SC-mViT as a promising deep learning solution for low-dose CT denoising, thereby warranting further validation in larger datasets.