The Unseen Aneurysm: Can Artificial Intelligence Bridge the Diagnostic Gap in Angiogram-Negative Subarachnoid Hemorrhage?

Few clinical scenarios carry as much diagnostic urgency and uncertainty as the patient who presents with a clear aneurysmal pattern of subarachnoid hemorrhage (SAH) but no identifiable culprit on initial vascular imaging. Yet, mortality in patients with an untreated ruptured aneurysm approaches 65% during the first year, and rebleeding rates reach 50% within 6 months, with the majority occurring in the first 24 to 48 hours. ¹ Today, up to 15% of patients with spontaneous SAH have no structural cause identified on initial computed tomography angiography (CTA), leaving clinicians to navigate a complex and often anxiety-laden decision tree of repeat imaging, advanced techniques, and watchful waiting. The review by Lindgren et al published in the Canadian Association of Radiologists Journal presents a timely and comprehensive imaging algorithm for precisely this patient population. ² The authors navigate the complexities of repeat digital subtraction angiography (DSA), vessel-wall MR imaging, and 7T imaging, ultimately proposing a patient-centered algorithm that accounts for bleeding pattern, clinical trajectory, and the strengths and limitations of each modality. What the review implicitly raises, is where does artificial intelligence (AI) fit into this diagnostic chain, and can it help us see what we currently cannot?

The diagnostic challenge in angiogram-negative SAH is a problem of perception and pattern recognition operating at the limits of human visual capacity. Blood blister-like aneurysms are flat, broad-based, and blend into the parent vessel wall. Perforating artery aneurysms arise from sub-millimeter branches that tax the spatial resolution of even state-of-the-art DSA. ³ Dissecting aneurysms may manifest only as subtle luminal irregularities or transient vessel wall changes that resolve between imaging sessions. Thrombosed aneurysms, by definition, are invisible to lumen-based techniques. In each of these scenarios, the aneurysm exists but escapes detection not because of inadequate technology per se, but because the signal is buried in noise, distorted by artifact, or camouflaged by normal anatomical variation. ⁴ This is precisely the domain in which AI could excel, if training data, validation standards, and clinical integration pathways are adequately addressed.

Modern deep learning architectures, particularly convolutional neural networks and transformer-based models, have demonstrated capacity to detect intracranial aneurysms on CT and MR angiography, in some studies matching or exceeding radiologist sensitivity for aneurysms larger than 3 mm. However, the frontier that matters most for angiogram-negative SAH is not the detection of obvious aneurysms but the identification of subtle, atypical, or partially obscured lesions that human readers miss on first and even second review. AI systems trained on large, annotated datasets that include these rare subtypes could function as a safety net, flagging regions of suspicion on CTA or DSA that warrant closer scrutiny or targeted advanced imaging. The key paradigm shift is moving AI from a confirmatory tool to an exploratory one, that is, from finding what we expect to revealing what we have not yet considered.

Beyond lesion detection, AI can contribute to the upstream decision of whether further imaging is warranted at all. The bleeding pattern on initial CT-scan is the single most important determinant of subsequent imaging strategy, yet its classification remains subjective and operator-dependent. Automated hemorrhage pattern classification using deep learning could standardize this critical triage step, reliably distinguishing perimesencephalic from aneurysmal distributions and identifying atypical patterns that mandate aggressive work-up. Similarly, AI-driven analysis of CTA source data, including vessel wall morphology, calcium scoring, and hemodynamic modeling, could generate individualized risk profiles that inform the timing and modality of repeat imaging rather than relying on fixed protocols.

Perhaps, AI may help us address what we do not know and what we do not see. Generative models and anomaly detection algorithms can be trained not on specific pathologies but on the statistical distribution of normal vascular anatomy. Deviations from this learned normality, even those too subtle for categorical diagnosis, could be flagged for longitudinal surveillance or advanced imaging with vessel-wall MRI or 7T. This unsupervised approach has the potential to capture entirely novel or unrecognized pathologies that fall outside existing classification schemes.

However, as AI models trained predominantly on saccular aneurysms, it may exhibit diminished sensitivity for the very lesions that matter most in angiogram-negative SAH: the dissections, blister aneurysms, and perforator lesions that are underrepresented in training datasets. Algorithmic confidence in the absence of disease can be dangerously reassuring if the model has never adequately learned to recognize these entities. Moreover, regulatory pathways for deploying AI tools in this specific acute neurovascular setting remain undefined, and the medicolegal implications of algorithmic false negatives in a life-threatening condition have yet to be addressed. Addressing this requires deliberate curation of training data enriched with rare subtypes, federated learning across institutions, and rigorous external validation in the specific population of initially negative SAH patients.

The imaging algorithm proposed in this review represents the current evidence-based standard. It is thorough, logical, and grounded in decades of accumulated clinical experience. But it is also, by necessity, a protocol designed for the average patient navigated by the average reader. Artificial intelligence offers the possibility of personalizing this algorithm in real time, sharpening human perception where it reaches its limits and measuring uncertainty where it remains hidden, and ultimately transforming angiogram-negative SAH from a diagnosis of exclusion into a diagnosis of precision. The unseen aneurysm need not remain unseen.