Author: Gary Y. Ge, Abdullah-Al-Zubaer Imran, Kazi Ramisa Rifa, Charles Mike Weaver, Jie Zhang 👨🔬
Affiliation: University of Kentucky 🌍
Purpose: With renewed attention on CT radiation dose management following CMS approval of new dose measures, establishing image quality–based target doses for every protocol has become essential. While global noise is adopted by CMS, its adequacy as a representative metric for image quality assessment (IQA) remains uncertain. This study evaluates the relationships between deep learning (DL)-based IQA scores and global noise metrics.
Methods: CT images from 105 chest scans (regular dose), 100 urogram scans (high dose), 100 abdominal scans (regular dose), and 100 renal stone scans (low dose) were retrospectively analyzed. IQA scores were calculated using the EfficientNet V2-L model, which achieved 3rd place in the Low-dose CT Perceptual IQA Grand Challenge 2023 and aligns closely with radiologists’ evaluations. Global noise metrics were calculated using Duke and Wisconsin methods. Linear regression and Pearson correlation coefficients (PCCs) quantified relationships between IQA scores and noise metrics.
Results: A strong negative correlation between IQA scores and global noise was observed for chest scans (PCCs: r = -0.86 to -0.82, R² = 0.74–0.67) and renal stone scans (PCCs: r = -0.89 to -0.85, R² = 0.79–0.72). Abdominal scans exhibited a moderate negative correlation (PCCs: r = -0.58 to -0.52, R² = 0.34–0.27), while urogram scans showed negligible to very weak correlations (PCCs: r = -0.21 to -0.18, R² = 0.04–0.03). The findings indicate that as global noise increases, IQA scores decrease, with the strength of this relationship varying by scan type. Chest and renal stone scans demonstrated the strongest trends, while urogram scans exhibited the weakest.
Conclusion: Global noise metrics strongly correlate with DL-based IQA scores at low doses but become less reliable at higher doses. These findings highlight the limitations of global noise in establishing image quality–based target doses, particularly for high-dose protocols.