Dual-Energy CT Boosts AI Lung Nodule Detection Accuracy
In a pivotal advance for early lung cancer screening, researchers in China have demonstrated that dual-energy computed tomography (CT) significantly enhances the performance of artificial intelligence (AI) systems in detecting pulmonary nodules—particularly small and subsolid lesions that are often missed in conventional imaging. The findings, published in China Medical Devices, suggest that integrating dual-energy CT protocols into routine clinical workflows could improve diagnostic sensitivity while simultaneously reducing patient radiation exposure.
Lung cancer remains the leading cause of cancer-related deaths globally, with China accounting for a substantial portion of new cases and fatalities each year. Early detection through low-dose CT screening has been shown to improve five-year survival rates, yet challenges persist in identifying subtle nodules, especially those under 4 millimeters in diameter or exhibiting subsolid characteristics. These lesions are clinically significant but notoriously difficult to detect, even for experienced radiologists.
Enter AI: deep learning algorithms trained on vast datasets have increasingly been deployed to assist in nodule detection, offering speed, consistency, and scalability. However, the efficacy of these systems is heavily dependent on image quality. Until now, most AI validation studies have relied on standard single-energy 120 kVp CT scans—the clinical norm for decades. The new study, led by Song Dongdong and colleagues at the Affiliated Zhongshan Hospital of Dalian University in collaboration with Infervision Technology Co., Ltd., challenges this status quo by rigorously evaluating how dual-energy CT fusion images influence AI performance.
The team enrolled 381 patients undergoing lung cancer screening between December 2018 and February 2019. Participants were randomly assigned to one of two scanning protocols: Group A (n=183) received conventional single-source 120 kVp scans, while Group B (n=198) underwent dual-energy scanning using a Siemens Somatom Definition Flash dual-source CT system with tube voltages of 100 kVp and Sn140 kVp. Both groups utilized CareDose 4D automatic exposure control to optimize radiation dose. From the dual-energy data, a synthetic 120 kVp fusion image was generated using a 4:6 weighting of the low- and high-energy acquisitions—a technique designed to combine high contrast from the 100 kVp dataset with low noise from the Sn140 kVp dataset.
All images were reconstructed at 1 mm slice thickness using a bone algorithm, then analyzed by a commercially available AI lung nodule detection software (InferRead CT Lung Research, Infervision). The algorithm, trained on over 400,000 annotated CT scans from multiple Chinese tertiary hospitals, automatically flagged suspected nodules and classified them by size (≥4 mm vs. <4 mm) and type (solid vs. subsolid).
To establish a reliable ground truth, three radiologists—two with over a decade of experience and one senior chief physician—jointly reviewed all scans according to the Chinese Expert Consensus on Diagnosis and Management of Pulmonary Nodules (2018). Discrepancies were resolved through consensus, yielding a definitive “gold standard” nodule set against which AI performance was measured.
The results were striking. In Group B (dual-energy fusion), the AI system achieved a nodule detection sensitivity of 91.5%, significantly higher than the 82.8% observed in Group A (single-energy), with a p-value of <0.001. This improvement held across all subcategories: for nodules ≥4 mm, sensitivity rose from 91.1% to 97.9%; for sub-4 mm nodules—a critical diagnostic blind spot—it jumped from 80.5% to 89.7%. Even more notably, detection of subsolid nodules, which carry a higher risk of malignancy, improved dramatically from 70.9% to 89.3%.
Crucially, this gain in sensitivity did not come at the cost of increased false alarms. The overall false positive rate per scan was slightly lower in Group B (1.4/CT) compared to Group A (1.6/CT). For solid nodules ≥4 mm, the false positive rate dropped from 1.0/CT to 0.8/CT. While the difference in false positives for subsolid nodules was not statistically significant, the substantial increase in true positives suggests a net clinical benefit—fewer missed cancers without overwhelming radiologists with spurious alerts.
Perhaps equally important, the dual-energy protocol delivered these advantages while reducing radiation exposure. The effective dose (ED) in Group B averaged 3.2 mSv, compared to 4.0 mSv in Group A—a 20% reduction that is both statistically and clinically meaningful, especially in screening populations that may undergo repeated scans over time. Dose-length product (DLP) and volumetric CT dose index (CTDIvol) also showed significant decreases in the dual-energy group (p < 0.001 for both).
These findings challenge assumptions about the neutrality of imaging protocols in AI performance. A recent study published in Radiology: Artificial Intelligence suggested that CT scanner manufacturer and radiation dose have minimal impact on deep learning model accuracy. Yet the Dalian team’s data indicate that image quality—shaped by acquisition physics—matters profoundly. The dual-energy fusion image, by blending the high contrast of 100 kVp with the low noise of Sn140 kVp (achieved via tin filtration that hardens the X-ray beam), yields a superior contrast-to-noise ratio. This enhanced image fidelity appears to provide the AI algorithm with clearer visual cues, enabling more reliable identification of faint or ambiguous nodules.
The implications extend beyond technical performance. In real-world clinical settings, even small improvements in sensitivity can translate into lives saved. Sub-4 mm and subsolid nodules are precisely the types most likely to be overlooked during routine reads, yet they may represent early-stage adenocarcinomas or pre-invasive lesions. By catching more of these at the screening stage, dual-energy CT paired with AI could shift diagnoses to earlier, more treatable phases—aligning with China’s national strategy to curb its soaring cancer burden through precision prevention and early intervention.
Moreover, the reduced radiation dose addresses a longstanding concern about the safety of population-wide CT screening. While low-dose protocols have mitigated risk, any further reduction enhances the risk-benefit calculus, potentially expanding eligibility to younger or higher-risk cohorts.
Despite its strengths, the study acknowledges limitations. The patient cohort, though randomized, was drawn from a single center and may not fully represent broader demographic or disease spectra. Future multicenter trials with larger, more diverse populations are needed to validate these results. Additionally, the study focused on detection—not characterization or malignancy prediction—so the impact on downstream diagnostic pathways remains to be explored.
Nevertheless, the evidence is compelling enough to prompt reconsideration of standard imaging protocols in AI-augmented radiology. As healthcare systems worldwide integrate AI into diagnostic workflows, the choice of acquisition technique can no longer be viewed as merely a hardware or dose optimization issue—it is a critical determinant of algorithmic efficacy.
For hospitals investing in next-generation CT infrastructure, dual-source systems capable of dual-energy imaging may offer more than advanced material decomposition or virtual non-contrast capabilities; they may serve as the optimal foundation for AI-driven precision diagnostics. In China, where both lung cancer incidence and AI adoption in healthcare are rising rapidly, this synergy could accelerate the transition from reactive treatment to proactive, data-informed prevention.
The research also underscores a broader principle in the age of medical AI: algorithmic performance is not absolute but context-dependent. Just as a self-driving car performs better on well-marked highways than on rural dirt roads, an AI diagnostic tool thrives on high-fidelity inputs. Optimizing the entire imaging chain—from X-ray generation to reconstruction—is essential to unlocking AI’s full potential.
As regulatory bodies and payers increasingly demand evidence of real-world clinical utility for AI tools, studies like this provide a roadmap: not just validating software in isolation, but evaluating it within optimized, integrated clinical pathways. The future of radiology may lie not in choosing between human expertise and machine intelligence, but in engineering the imaging environment where both can perform at their best.
In conclusion, the integration of dual-energy CT fusion imaging with AI-powered nodule detection represents a significant step forward in lung cancer screening. It delivers higher sensitivity, lower false positives, and reduced radiation—all critical metrics for scalable, safe, and effective population health programs. As this approach matures, it could set a new benchmark for early cancer detection worldwide.
Song Dongdong¹, Zhu Xiaoming¹, Zhu Lijuan¹, Gu Jun², Wu Jianlin¹, Zhang Qing¹
¹Department of Radiology, Affiliated Zhongshan Hospital of Dalian University, Dalian, Liaoning 116001, China
²Institute of Global Clinical Research Collaboration, Infervision Technology Co., Ltd., Beijing 100025, China
China Medical Devices, Vol. 36, No. 02, 2021
DOI: 10.3969/j.issn.1674-1633.2021.02.018