Deep learning for cardiac CT segmentation for congenital heart disease: A systematic review.
Sinling Tiffany Yu, Zahra Amini, Shereen Fouad et al.
✦ AI-curated · Sources linked
In 30 seconds
This systematic review evaluates deep learning (DL) approaches for cardiac CT segmentation in congenital heart disease (CHD), analyzing 14 studies. The UNet architecture was most prevalent, with the highest Dice Similarity Coefficient (DSC) of 0.945 for aorta segmentation using a hybrid ResNet model. The review highlights challenges in clinical implementation and reporting quality.
Key findings
- The UNet architecture and its variants were used in 8 out of 14 studies.
- The highest reported Dice Similarity Coefficient for cardiac structures was 0.945 for aorta segmentation.
- Only two studies performed external validation, indicating a gap in clinical applicability.
- Small vascular structures consistently underperformed compared to whole heart segmentation.
Why it matters
Understanding the capabilities and limitations of deep learning in cardiac CT segmentation for CHD is crucial for clinicians. Improved segmentation can enhance diagnostic accuracy and treatment planning, but challenges in data quality and reporting must be addressed for effective clinical use.