AI and Advanced CT Imaging Team Up to Spot Brain Aneurysms with Greater Precision
In a significant leap forward for diagnostic radiology, researchers from Chongqing Medical University have demonstrated that combining cutting-edge artificial intelligence (AI) with virtual monoenergetic imaging technology can dramatically improve the detection of intracranial aneurysms—potentially life-threatening bulges in brain arteries. Their findings, published in the peer-reviewed journal Chinese Medical Devices, offer clinicians a powerful new tool to enhance accuracy, reduce diagnostic errors, and ultimately save lives.
The study, led by Dr. Wei Xin and colleagues from the Department of Radiology at The Second Affiliated Hospital of Chongqing Medical University, focused on optimizing head and neck CT angiography (CTA)—a non-invasive imaging technique widely used to visualize blood vessels in the brain and neck. While CTA is already considered highly sensitive for detecting aneurysms, its effectiveness can be hampered by image artifacts caused by dense skull bones or suboptimal contrast resolution. This is where dual-energy CT (DECT) and AI come into play.
Dual-energy CT works by acquiring images using two different X-ray energy levels simultaneously. This allows radiologists to generate “virtual” single-energy images across a wide spectrum—from 40 keV to 120 keV—that simulate what the scan would look like if it had been performed at just one specific energy level. The key advantage? By selecting the optimal energy level, clinicians can enhance the contrast between blood vessels and surrounding tissues, effectively “tuning” the image for maximum clarity.
But even with improved image quality, human interpretation remains subjective and prone to fatigue or oversight—especially when dealing with complex vascular anatomy. Enter artificial intelligence. The research team employed an AI-assisted diagnostic platform called Cerebral Doc, which uses deep learning algorithms to automatically analyze CTA scans and flag potential aneurysms. The system was trained to segment vessels, remove bone structures, and identify suspicious areas based on patterns learned from prior cases.
What sets this study apart is its rigorous methodology. The researchers retrospectively analyzed data from 41 patients who had undergone head and neck CTA between August and September 2020 due to suspected intracranial aneurysms. All scans were reconstructed using both standard linear fusion (F_0.6, equivalent to a conventional 120 kV scan) and nine virtual monoenergetic image sets ranging from 40 to 120 keV in 10 keV increments. Each set was then processed through the AI system.
To establish a gold standard for comparison, two senior radiologists independently reviewed the F_0.6 images using a double-blind protocol. Any discrepancies were resolved through consensus, and the final diagnosis was verified by a deputy chief physician with over a decade of experience in neuro-radiology. This meticulous approach ensured that the AI’s performance could be evaluated against expert human judgment—a critical component of scientific validity.
The results were compelling. When evaluating diagnostic accuracy—the proportion of true positives among all actual cases—the AI system performed best at 50 keV, achieving a remarkable 78% accuracy rate. At this energy level, the area under the receiver operating characteristic (ROC) curve reached 0.796, indicating strong discriminatory power. More importantly, there was no statistically significant difference between the AI’s performance at 50 keV and that of the senior radiologists, suggesting that the AI can match human expertise under optimal imaging conditions.
Interestingly, performance varied significantly across energy levels. At lower energies like 40 keV, while image contrast was highest, the AI’s accuracy dropped to just 51%. Similarly, at higher energies such as 110 and 120 keV, accuracy fell to 64% and 60%, respectively—both statistically inferior to human readers. However, within the range of 50 to 100 keV, the AI consistently matched or closely approximated human diagnostic capability, with accuracies hovering around 70–78%.
This finding underscores a crucial point: not all virtual monoenergetic reconstructions are created equal. While low-energy images provide superior contrast-to-noise ratios (CNR), they may also introduce excessive noise or artifacts that confuse AI models. Conversely, high-energy images reduce noise but sacrifice contrast, making subtle lesions harder to detect. The sweet spot appears to lie around 50 keV, where the balance between signal enhancement and image fidelity enables the AI to function optimally.
Quantitatively, the team measured several objective parameters including CT attenuation values, signal-to-noise ratio (SNR), and CNR at the common carotid artery bifurcation—a standardized anatomical landmark. They found that compared to standard F_0.6 images, 50 keV reconstructions showed a 92.47% increase in CT value, an 18.65% improvement in SNR, and an 11.87% boost in CNR. These gains directly translate to clearer vessel delineation and better lesion conspicuity—factors that are essential for accurate diagnosis.
Moreover, the study revealed that certain technical factors can influence AI performance. For instance, overly high CT values (as seen at 40 keV) or excessively low values (at 110+ keV) can lead to misinterpretations. This suggests that protocols involving contrast injection rates, bolus timing, and reconstruction settings must be carefully calibrated to support AI-based diagnostics. It also opens up opportunities for future refinement—for example, adjusting neural network weights or training datasets to accommodate extreme energy levels.
One of the most intriguing aspects of the research is how the AI system identifies false positives and negatives. According to the authors, misdiagnoses often occurred when the algorithm mistook normal vascular dilations or large plaque formations for aneurysms. Missed diagnoses typically involved very small aneurysms or cases where aggressive segmentation inadvertently removed portions of the vessel wall along with adjacent bone. These insights highlight the importance of integrating clinical context and anatomical knowledge into AI decision-making processes—not just relying on pixel-level pattern recognition.
Despite these promising outcomes, the researchers acknowledge limitations. First, the study did not use digital subtraction angiography (DSA)—the current clinical gold standard for diagnosing cerebral aneurysms—as a reference. Second, the sample size of 41 patients, though sufficient for initial validation, is relatively small for generalizing results across diverse populations. Third, the AI system exhibited sensitivity to patient positioning; minor shifts in orientation sometimes led to incorrect vessel labeling. Finally, the dataset was drawn exclusively from a single institution, potentially limiting external validity.
Nevertheless, the implications of this work are profound. Intracranial aneurysms affect approximately 3–5% of the general population, with peak incidence occurring between ages 55 and 60. Many remain asymptomatic until rupture, which carries a mortality rate of up to 50%. Early detection via non-invasive imaging like CTA is therefore vital. By augmenting radiologists with AI tools that perform reliably across multiple energy levels, healthcare systems can streamline workflows, reduce turnaround times, and minimize missed diagnoses—particularly in busy emergency departments or rural clinics lacking subspecialist expertise.
From a technological standpoint, this study exemplifies the convergence of advanced imaging hardware and intelligent software. Dual-energy CT scanners, once considered niche equipment, are becoming increasingly common in major hospitals worldwide. Meanwhile, AI platforms are evolving beyond simple classification tasks toward comprehensive diagnostic assistants capable of multi-step reasoning and contextual awareness. Together, they represent a paradigm shift in medical imaging—one that prioritizes precision, efficiency, and accessibility.
Looking ahead, the research team plans to expand their investigation by incorporating DSA-confirmed cases, evaluating AI performance across different aneurysm morphologies and sizes, and testing the system’s robustness across various scanner manufacturers and reconstruction algorithms. They also aim to develop adaptive AI models that can self-optimize based on individual patient characteristics or institutional preferences.
For practicing radiologists, the message is clear: AI is not here to replace you—it’s here to empower you. When paired with optimized imaging protocols like VMI+ at 50 keV, AI can serve as a reliable second reader, catching subtle abnormalities that might otherwise slip through the cracks. In time, such systems may become integral components of routine clinical practice, helping to standardize care, reduce variability, and elevate diagnostic confidence across the board.
In conclusion, the integration of virtual monoenergetic imaging and artificial intelligence marks a transformative step in the fight against intracranial aneurysms. As technology continues to evolve, so too will our ability to detect, diagnose, and treat cerebrovascular disease with unprecedented speed and accuracy. The path forward lies not in choosing between machines and humans, but in forging synergistic partnerships that leverage the strengths of both.
Wei Xin, Chen Weijuan, Deng Hao, Yu Han, Cao Wenting, Chen Jinhua, Liu Yang. Research of Dual Energy CT VMI+ Technology Combined with Artificial Intelligence-Assisted Diagnosis System in the Diagnosis of Intracranial Aneurysm. Chinese Medical Devices, 2021, Vol.36 No.10, pp.100–107. doi:10.3969/j.issn.1674-1633.2021.10.023