Qualitative and quantitative assessment of deep learning-based raw data motion correction in atrial fibrillation: benefit in high heart rate and motion-prone segments
H Kawai, M Sarai, N Uchida, H Takahashi, Y Motoike, T Muramatsu, M Harada, S Motoyama, H IzawaAbstract
Background
Motion artifacts remain a major challenge in coronary computed tomography angiography (CTA) for patients with atrial fibrillation (AF). Recently, novel motion correction (MC) algorithms based on deep learning of raw data have become available, potentially overcoming the limitations of conventional reconstruction in irregular and rapid heart rhythms.
Purpose
We aimed to evaluate the clinical benefit of this deep learning-based raw data MC specifically within an AF population, focusing on its impact on image quality across different heart rates and coronary segments.
Methods
We retrospectively analysed 100 patients with AF who underwent coronary CTA using a 320-row scanner. Images were reconstructed with and without a novel deep learning-based raw data MC algorithm (MC(+) and MC(-), respectively). For MC(+), the deep learning algorithm estimated the motion vector field from raw data, enabling high-quality image generation using only one beat for half-reconstruction. For MC(-), the better image quality from either conventional half or segment reconstruction was selected for comparison. Quantitative analysis was performed on the middle segments of the right coronary artery (RCA) and left circumflex artery, as these typically run perpendicular to the axial plane, allowing for direct cross-sectional evaluation. After excluding segments with atherosclerotic plaque, 48 patients were included in the evaluation of luminal boundary sharpness using edge rise distance (ERD) and edge rise slope (ERS). For qualitative analysis, two independent observers graded image quality for all coronary segments of all 100 patients using a 5-point scale.
Results
Luminal boundary sharpness significantly improved in the MC(+) group compared with the MC(-) group (reduced ERD and increased ERS, both p < 0.001). Qualitative scores were also significantly higher in the MC(+) group than in the MC(-) group (4.1±0.3 vs. 3.8±0.5, p < 0.001). This improvement was significantly more pronounced in patients with higher heart rates (p for interaction < 0.001). In this overall assessment, the improvement in image quality was most notable in the middle segment of the RCA among all coronary segments.
Conclusions
In patients with AF, deep learning-based raw data MC significantly enhances coronary image quality. This is the first report to demonstrate that these benefits are particularly evident in patients with higher heart rates and in coronary segments prone to motion, suggesting that this technology provides more robust diagnostic performance for coronary CTA in the challenging AF population.