A patient presents to the emergency department with a routine abdominal pain. The emergency CT scan shows nothing obvious. He goes home, and returns a few hours later in cardiac arrest. Another patient, followed in oncology for metastatic cancer, undergoes a weekly surveillance scan. The incidental pulmonary embolism visible on his images will go unnoticed for several days, long enough for the radiologist to work through the stack of pending reports.
These two scenarios are neither fiction nor negligence: they illustrate two well-documented blind spots in emergency radiology. Acute aortic syndrome (AAS), which encompasses aortic dissection, intramural hematoma, and penetrating atherosclerotic ulcer, carries a mortality rate of 40 to 50% within the first 48 hours without treatment, with progression of 1 to 2% per hour. Incidental pulmonary embolism (IPE) in cancer patients, meanwhile, multiplies the risk of mortality by 4 to 8 compared to the general population, and yet remains underdiagnosed in the majority of care centers. In the face of these silent emergencies, artificial intelligence (AI) is progressively establishing itself as a proactive detection tool.
Detecting Acute Aortic Syndrome on Non-Contrast CT
Contrast-enhanced CT is the absolute reference standard for diagnosing acute aortic syndrome. Yet in many resource-limited settings, as well as in emergency contexts where injection is contraindicated or delayed, non-contrast CT remains the available initial examination. Its diagnostic sensitivity for AAS, when interpreted alone by a radiologist, is notoriously low, representing a major clinical challenge for a condition where every hour counts.
AI algorithms reach a sensitivity of 91 to 94% and a specificity of 99%, with a negative predictive value approaching 100%. In other words, a negative AI result constitutes a very strong argument for ruling out AAS in a patient whose initial clinical presentation points toward another cause.
The interpretability of these systems constitutes another decisive advantage. Rather than functioning as a “black box” that returns a probability score without justification, the most advanced algorithms generate activation maps indicating to the radiologist the precise anatomical zones that triggered the alert. This transparency reinforces clinical confidence in the decision.
Incidental Pulmonary Embolism in Oncology: AI as a Safety Net
In oncology, the context is fundamentally different: patients do not present with acute chest pain. They undergo scheduled surveillance scans, often thoraco-abdomino-pelvic (TAP) protocols, to evaluate their response to treatment. Incidental pulmonary embolism is not being looked for. It is discovered, or not, on the margins of the main analysis.
Yet cancer patients carry a thromboembolic risk 12 times higher than the general population, and up to 23 times higher when receiving active chemotherapy. The sensitivity of radiologists for detecting these incidental emboli varies, a concerning variation for a condition whose rapid management significantly reduces mortality.
The value of AI here lies in its capacity to sort the radiologist’s worklist in real time, immediately bringing positive cases to the top of the stack.
Two Pathologies, One Logic: AI as an Early Warning System in the Radiology Workflow
What is striking when reading this data is the convergence of integration models. In both cases, AAS on non-contrast CT and IPE on oncology CT, the added value of AI does not lie in replacing the radiologist, but in a clearly defined function: early alerting and intelligent worklist triage.
Both systems operate according to a similar paradigm. The algorithm analyzes images in real time as they arrive in the picture archiving and communication system (PACS). If a pathology is detected, an alert is sent to the radiologist via a pop-up notification, indicating that this examination should be prioritized. Negative cases, whose high negative predictive value guarantees the near-certain absence of urgent pathology, remain in the normal queue.
This model addresses several structural constraints of contemporary radiology: the continuous growth of examination volumes, reading fatigue during prolonged on-call periods, the absence of specialized radiological coverage in smaller institutions and during off-peak hours, and the natural tendency to prioritize the main analysis (oncological surveillance, trauma workup) at the expense of urgent incidental findings.
One point deserves to be underlined: the complementarity between AI and the clinician works in both directions. If AI detects cases missed by radiologists, radiologists in turn correct errors made by the algorithm. In comparative studies, cases misclassified by AI are often correctly identified by physicians, and vice versa. This symmetry of blind spots argues for a systematically combined approach, in which neither actor works in isolation.
Current Limitations and Points of Vigilance for Responsible Integration
While published performance figures are promising, informed deployment requires keeping a clear-eyed view of several persistent limitations.
Aortic anatomical variability between populations, as well as differences in acquisition protocols across equipment and institutions, demand vigilance when institutions select their solution.
False positives represent a real organizational constraint, particularly for IPE detection algorithms. A high false-positive rate does not compromise patient safety, but generates additional verification work for the radiologist, which can be counterproductive if it leads to alert desensitization. Minimizing false positives is therefore a priority industrial challenge.
The pathological coverage of algorithms remains incomplete. For AAS, penetrating atherosclerotic ulcers (PAU) are systematically less well detected than aortic dissections, due to their more subtle morphology. For IPE, subsegmental emboli of uncertain clinical significance raise unresolved questions about the appropriate clinical course of action. These gaps serve as a reminder that the overall performance of an algorithm does not remove the need for detailed subgroup-level analysis.
Finally, the phenomenon of excessive deference to the algorithm, described in pediatric radiology as well as in emergency medicine, applies fully to this context. A physician who abandons a correct diagnosis because AI did not confirm it commits a potentially fatal error. Training users in the critical interpretation of algorithmic outputs is not a luxury: it is a prerequisite for patient safety.
Toward Augmented Emergency Radiology
AI applied to the detection of cardiovascular and thromboembolic emergencies in imaging is no longer a research object: it is a clinically validated tool whose benefits can be measured in time gained on diagnosis and lives potentially saved. Its high negative predictive value makes it a powerful safety net for patients managed in environments with limited radiological coverage.
The challenge ahead lies in extending pathological coverage: algorithms capable of simultaneously detecting multiple life-threatening emergencies on a single examination, with homogeneous performance across the full spectrum, including aortic dissection, pulmonary embolism, intracranial hemorrhage, tamponade, and pneumothorax, would constitute a genuine revolution in automated radiological triage.
In clinical practice, the challenge is organizational and cultural: defining clear integration protocols, training teams in the critical interpretation of algorithmic alerts, and designing feedback loops that allow algorithms to continuously learn from errors observed in real-world practice.
Sources
Hu Y, Xiang Y, Zhou YJ, et al. AI-based diagnosis of acute aortic syndrome from noncontrast CT. Nat Med. 2025 Nov;31(11):3832-3844. doi: 10.1038/s41591-025-03916-z. PMID: 40835970. https://pubmed.ncbi.nlm.nih.gov/40835970/
Ammari S, Camez AO, Ayobi A, et al. Contribution of an Artificial Intelligence Tool in the Detection of Incidental Pulmonary Embolism on Oncology Assessment Scans. Life (Basel). 2024 Oct 22;14(11):1347. doi: 10.3390/life14111347. PMID: 39598146. https://pubmed.ncbi.nlm.nih.gov/39598146/
