Revolutionising acute aortic syndrome diagnosis: The role of artificial intelligence in non‐contrast computed tomography

Rapid and accurate diagnosis of acute aortic syndrome (AAS) is crucial.1 Without treatment, 40%–50% of patients die within 48 hours, and mortality increases by 1%–2% per hour of delay.2, 3 Therefore, any delay in diagnosis substantially worsens the prognosis.1 However, AAS is difficult to diagnose because its symptoms are nonspecific4 and routine tests are unreliable.5 Although aortic computed tomography angiography (CTA) is considered the gold standard for diagnosing AAS,1 it is expensive, carries risks such as anaphylaxis and nephrotoxicity from contrast agents,5 and is often not available in resource-limited settings like China6 and other low- and middle-income countries.7 Correct diagnosis of AAS currently relies heavily on clinical suspicion, as there is no simple, dedicated diagnostic algorithm.1 Noncontrast CT is a convenient, cost-effective, and widely used imaging modality, and its potential for screening AAS has been explored previously.8 However, it still lacks adequate sensitivity and specificity for accurate AAS diagnosis.9 In this context, artificial intelligence (AI) applied to non-contrast CT scans offers a promising approach. Here, we developed an AI-based warning system called iAorta.10 The system is based on a deep learning model trained on 3350 aortic CTA scans. It uses non-contrast phase images supervised by diagnostic labels and lesion segmentations transferred from arterial phase scans via image registration. The model detects acute aortic syndrome at the patient level, segments the aorta and true lumen precisely, and locates lesion regions with high reliability. To improve interpretability, the system produces activation maps that visually highlight areas corresponding to pathological regions, providing slice-level visual explanations for different disease subtypes (Figure 1a). Integrated into a browser-server platform, iAorta processes data in real time, sends alerts to radiologists, and supports clinical decision-making, improving both speed and accuracy of AAS detection in emergency settings. The performance of iAorta has been rigorously validated (Figure 1b). In a multicenter model validation across eight centres (n = 20,750), the model achieved a sensitivity of 0.954–0.984 and specificity of 0.929–0.947 for detecting AAS on non-contrast CT. In a large-scale real-world retrospective study (n = 137,525), the model showed consistent performance across different scanning protocols, with sensitivity of 0.913–0.942 and specificity of 0.991–0.993. In a reader study, the diagnostic sensitivity of medical trainees increased by 52.4% (from 0.603 to 0.919) with the model's assistance, reaching a level similar to that of specialist experts. During a pilot implementation at a Chinese hospital, iAorta was integrated into the routine workflow, automatically analysing CT images from the hospital PACS. It identified patients at risk of AAS and provided real-time alerts to radiologists via an interactive interface. Between 20 December 2024, and 20 February 2025, iAorta helped detect 21 out of 22 AAS cases among 15,584 consecutive emergency department patients who underwent non-contrast CT, missing only one penetrating atherosclerotic ulcer. The system demonstrated high diagnostic accuracy, with sensitivity of 0.955 (95% confidence interval [CI] 0.864–1.000) and specificity of 0.994 (95% CI 0.993–0.995). The average time to diagnosis for confirmed AAS cases was reduced to 102.1 min (range 75–133), while a study based on the International Registry of Acute Aortic Dissection reported the median time from arrival at the emergency department to diagnosis for Stanford type A aortic dissection, even with typical symptoms, is 4.3 h.11 Thus, iAorta contributes to a substantial reduction in diagnostic delay. (See Figure 1b). While the potential of using non-contrast CT for screening AAS has been explored previously,8 the diagnostic performance of non-contrast CT alone has not been well characterised. iAorta changes the screening approach for AAS by making non-contrast CT a primary, efficient, and reliable clinical decision tool rather than only an auxiliary imaging method. For many patients with atypical chest pain, iAorta allows clinicians to make triage decisions using the most readily available initial imaging information. It helps direct emergency CTA resources to high-risk patients while providing low-risk patients with a reliable exclusion based on its high negative predictive value, reducing unnecessary invasive procedures and related risks (Figure 2a). iAorta also provides radiologists and emergency physicians with clear, verifiable information to support clinical judgment through visualisation outputs. This transparency allows the AI to act as a decision-support tool whose reasoning can be understood and checked, rather than a “black box”. By supporting collaboration between clinical expertise and AI analysis, iAorta encourages long-term integration into routine practice, addressing a key barrier to sustainable AI use in medicine12 (Figure 2b). What's more, iAorta addresses the critical “last mile” of AI adoption by serving as a comprehensive equaliser across both healthcare infrastructure and clinical expertise.13 It ensures accessibility and scalability by using widely available non-contrast CT equipment. It allows advanced diagnostic capabilities to extend from top-tier academic centres to resource-limited community hospitals, effectively mitigating regional disparities in healthcare access. By minimising inter-observer variability14 and the reliance on individual experience, the system ensures consistent, expert-level accuracy for high-stakes acute conditions, optimising patient management. Looking ahead, iAorta establishes a reproducible framework for unlocking the latent diagnostic value of non-contrast CT. This model serves as a blueprint for screening other time-sensitive conditions, such as non-ST-segment-elevation acute coronary syndrome, pulmonary embolism and oesophageal rupture,4 paving the way for a comprehensive “one-scan, multiple-screening” ecosystem.15 Eventually, this evolution transforms routine imaging from a specialised diagnostic aid into a generalised critical care platform, maximising existing infrastructure to revolutionise emergency diagnostic efficiency. This work was funded by grants from the National Natural Science Foundation of China (Grant No. 825B2011 to Yujian Hu) and the Technical Innovation key project of Zhejiang Province (2024C03023 to Hongkun Zhang). The authors declare no conflict of interest.