CT Scan Dose Does Not Affect AI Lung Nodule Detection Accuracy, Study Finds
A new study conducted by researchers at West China Hospital, Sichuan University has revealed that the accuracy of artificial intelligence (AI) systems in detecting lung nodules during computed tomography (CT) scans is not significantly influenced by radiation dose levels. Instead, the performance of these AI tools depends more heavily on the characteristics of the nodules themselves and the specific AI system used. The findings, published in CT Theory and Applications, challenge the assumption that higher radiation doses are necessary for reliable AI-assisted diagnosis and suggest that low-dose CT protocols can be safely used without compromising detection sensitivity.
Lung cancer remains the leading cause of cancer-related mortality worldwide, with early detection playing a critical role in improving survival outcomes. When identified at an early stage, the five-year survival rate for lung cancer can reach between 70% and 90%. In most cases, early-stage lung cancer presents as solitary pulmonary nodules—small, round abnormalities in the lung tissue that may be benign or malignant. The accurate identification and characterization of these nodules are therefore essential for timely intervention and effective treatment planning.
In recent years, artificial intelligence has emerged as a powerful tool in medical imaging, particularly in the field of radiology. AI algorithms, especially those based on deep learning, have demonstrated the ability to rapidly analyze large volumes of imaging data, flag potential abnormalities, and assist radiologists in making more accurate and efficient diagnoses. In chest CT screening, AI systems are increasingly being deployed to detect pulmonary nodules, reducing the likelihood of human oversight and improving workflow efficiency.
Despite the growing adoption of AI in clinical practice, there remains a lack of standardized guidelines regarding optimal imaging protocols for AI-assisted diagnosis. Various technical factors—such as tube voltage, tube current, reconstruction algorithms, image resolution, and slice thickness—can influence image quality and, potentially, the performance of AI systems. Among these, radiation dose, typically modulated through adjustments in tube voltage (kV) and tube current (mAs), has been a subject of particular interest due to its direct impact on patient safety.
The principle of ALARA (As Low As Reasonably Achievable) emphasizes minimizing radiation exposure while maintaining diagnostic image quality. However, reducing radiation dose can lead to increased image noise and reduced signal-to-noise ratio, which may affect the visibility of small or low-contrast lesions. This raises a critical question: Can AI systems maintain high detection accuracy under low-dose CT conditions? If so, it would support the use of ultra-low-dose protocols in population-based lung cancer screening programs, thereby reducing long-term radiation risks for patients.
To address this question, a team led by Deying Wen, a radiology technician at West China Hospital, designed a controlled phantom study to systematically evaluate the impact of CT scan dose on the efficacy of multiple commercial AI systems in detecting lung nodules. The study, co-authored by Xuelin Pan, Hui Yao, Jijie Li, Qiao Deng, Lu Tang, Xi Wu, and Jiayu Sun, was published in the August 2021 issue of CT Theory and Applications (DOI: 10.15953/j.1004-4140.2021.30.04.06).
The researchers used a standard adult male chest X-ray/CT phantom (Kyoto Kagaku PH-1), a physical model designed to simulate human thoracic anatomy at a 1:1 scale. Embedded within the phantom were 15 spherical simulated lung nodules of varying sizes (3, 5, 8, 10, and 12 mm in diameter) and densities (measured in Hounsfield Units, or HU). Three distinct density levels were represented: +100 HU (simulating solid nodules), -630 HU, and -800 HU (both simulating ground-glass or subsolid nodules). These nodules were randomly distributed within the lung fields to mimic real-world clinical scenarios.
The phantom was scanned 50 times using a uCT 780 scanner (United Imaging Healthcare, Shanghai), with variations in tube voltage (70, 80, 100, 120, and 140 kV) and tube current (ranging from 20 to 200 mAs in 20 mAs increments). All other scanning parameters, including scan range, reconstruction slice thickness (1 mm), and reconstruction interval (1 mm), were kept constant across all acquisitions. This rigorous control of variables ensured that any observed differences in AI performance could be attributed solely to changes in radiation dose.
After image acquisition, the dataset was analyzed using three commercially available AI systems from leading Chinese medical AI companies—referred to in the study as Company A, Company B, and Company C (corresponding to United Imaging Intelligence, Deepwise, and Infervision, respectively, though the authors emphasized that the labeling was alphabetical and not indicative of performance ranking). Each AI system was tasked with detecting and localizing all 15 nodules in every scan, and the results were recorded in terms of true positives, false negatives, and false positives.
The primary outcome measures were detection rate (equivalent to sensitivity), false negative rate (the proportion of nodules missed), and false positive rate (the number of non-nodule structures incorrectly flagged as nodules per scan). Statistical analyses were performed using Pearson’s chi-squared test or Fisher’s exact test for categorical data, and the Kruskal-Wallis H test for non-parametric comparisons of false positive rates. A significance level of α = 0.05 was used, with Bonferroni correction applied for multiple comparisons.
The results yielded several key insights. First, and most notably, there was no statistically significant difference in detection or false negative rates across different tube voltage or tube current settings for any of the three AI systems. This held true across all nodule sizes and densities, indicating that even at the lowest dose setting (70 kV and 20 mAs), the AI systems were able to detect nodules with the same sensitivity as at higher doses. This finding supports the feasibility of using ultra-low-dose CT protocols in conjunction with AI for lung nodule screening, aligning with the ALARA principle.
However, the study did find that tube voltage had a significant impact on false positive rates—the number of non-nodular structures incorrectly identified as nodules. For Company A, false positive rates were highest at lower kV settings (70 and 80 kV), with an average of 6.89 and 6.44 false positives per scan, respectively, compared to 3.89 at 140 kV. In contrast, Company B exhibited the lowest false positive rate at 70 kV (1.00 per scan) and the highest at 100 kV (2.00 per scan). Company C showed a similar pattern to Company A, with higher false positive rates at lower kV levels.
These divergent patterns suggest that different AI algorithms respond differently to changes in image noise and contrast, which vary with tube voltage. Lower kV settings produce higher contrast between soft tissues and air but also increase image noise, which may lead to over-detection of normal anatomical structures such as blood vessels or bronchial walls as nodules. The fact that Company B performed best at 70 kV indicates that its algorithm may be better optimized for low-dose, high-contrast imaging environments.
When comparing the three AI systems directly, the researchers found that Company B and Company C consistently outperformed Company A in overall detection accuracy, particularly for challenging nodules. For instance, all three systems struggled with 3 mm nodules, but Company B and C achieved higher detection rates than Company A. Similarly, nodules with a density of -800 HU—simulating very subtle ground-glass opacities—were more frequently missed by Company A (38.22% detection rate) compared to Company B (56.44%) and Company C (68.00%).
Interestingly, Company B showed superior performance specifically at 70 kV for certain nodule types. For +100 HU solid nodules, detection rate reached 100% at 70 kV, significantly higher than at 120 kV (80%) or 140 kV (80%). Similarly, for 3 mm nodules, detection rate was 33.33% at 70 kV, compared to 0% at both 120 kV and 140 kV. This suggests that for this particular AI system, lower kV settings enhance contrast resolution, making small or low-contrast nodules more visible.
In contrast, Company A performed best at 100 kV, where it achieved a balance between lower false positive rates and stable detection rates. Company C also showed optimal performance at 70 kV, though its false positive rate was higher than Company B’s at that setting.
Based on these findings, the authors concluded that while CT scan dose does not significantly affect the sensitivity of AI systems in detecting lung nodules, the choice of optimal scanning parameters should be tailored to the specific AI platform being used. They recommended 70 kV as the preferred tube voltage for Company B and Company C, and 100 kV for Company A, to maximize detection accuracy while minimizing false alarms.
The study also highlighted the importance of nodule characteristics in AI performance. As expected, larger nodules (8–12 mm) were detected with near-perfect accuracy across all systems, while 3 mm nodules posed a significant challenge. This is clinically relevant, as very small nodules are often considered indeterminate and may not require immediate intervention. The authors noted that if 3 mm nodules were excluded from analysis, the overall detection rates for the AI systems would increase substantially—reaching 82.41% for Company A, 89.63% for Company B, and 93.89% for Company C.
The implications of this research are significant for clinical practice. As lung cancer screening programs expand globally, particularly in high-risk populations, the integration of AI into routine CT interpretation is expected to grow. This study provides empirical evidence that low-dose CT protocols can be used safely and effectively with AI assistance, reducing patient radiation exposure without sacrificing diagnostic confidence.
Moreover, the findings underscore the need for AI vendors to optimize their algorithms for specific imaging conditions. Rather than assuming that one-size-fits-all, radiologists and imaging centers should consider the compatibility between their CT scanners, scanning protocols, and AI software. The study’s use of a physical phantom allows for reproducible testing, suggesting that similar evaluations could be conducted by hospitals to validate AI performance under their own operational settings.
The research team acknowledged several limitations. The use of a phantom, while advantageous for controlling variables, does not fully replicate the complexity of real human lungs, which may contain emphysema, fibrosis, or other pathologies that can obscure nodules. Additionally, the simulated nodules were perfectly spherical and homogeneous, whereas real nodules often have irregular shapes and heterogeneous internal structures. Future studies involving real patient data and diverse lung pathologies will be necessary to confirm these findings in clinical settings.
Nevertheless, this study represents a significant step toward evidence-based integration of AI in radiology. By systematically evaluating the interaction between imaging parameters and AI performance, the authors provide valuable guidance for optimizing lung nodule detection protocols. Their work reinforces the idea that AI is not a standalone diagnostic tool but a component of a larger imaging ecosystem, where scanner settings, reconstruction methods, and algorithm design must be carefully aligned to achieve the best outcomes.
As AI continues to evolve, such rigorous, independent evaluations will be essential to ensure that these technologies deliver on their promise of improving patient care. The study by Wen, Pan, Yao et al. serves as a model for how to approach the validation of AI in medical imaging—not just asking whether it works, but under what conditions it works best.
CT Scan Dose Does Not Affect AI Lung Nodule Detection Accuracy, Study Finds
Deying Wen, Xuelin Pan, Hui Yao, Jijie Li, Qiao Deng, Lu Tang, Xi Wu, Jiayu Sun, West China Hospital, Sichuan University
CT Theory and Applications, DOI: 10.15953/j.1004-4140.2021.30.04.06