Advancements in Machine Learning for Cerebral Aneurysm Detection: Progress at a Steady Pace

You never tell us in exact words how many strokes were prevented. 

Advancements in Machine Learning for Cerebral Aneurysm Detection: Progress at a Steady Pace

Nam Hoang Le and

Edgar A. Samaniego

Originally published14 Nov 2023https://doi.org/10.1161/SVIN.123.001122Stroke: Vascular and Interventional Neurology. 2023;3:e001122

• This article is a commentary on the following

We have read with great interest the article by Kim et al. entitled: “Machine‐Learning Enabled Detection of Unruptured Cerebral Aneurysms Improves Detection Rates and Clinical Care.” The authors explore the capabilities of a machine learning algorithm in the crucial task of detecting unruptured cerebral aneurysms (UCAs). To achieve this, the authors harness the power of the commercial software Viz ANEURYSM, a deep learning framework that has already demonstrated its effectiveness in a 5‐fold cross‐validation study.¹

The article presents a case study demonstrating the effectiveness of Viz ANEURYSM in assisting radiologists in identifying aneurysms in a large data set. It has been shown that deep learning is able to detect cerebral aneurysms in magnetic resonance angiography and computed tomography angiography (CTA), yielding commendable detection rates.^{2, 3, 4, 5, 6} Thus, with the support of artificial intelligence, the future of automatic detection of cerebral aneurysms is becoming a reality. However, to fully realize this potential, it is imperative that further studies be conducted to validate the practical utility of these innovative tools, ensuring their accuracy and efficacy in the detection of cerebral aneurysms as promised.

The authors report a positive predictive value of 62% in detecting UCAs. Viz ANEURYSM flagged 50 CTAs, and 31 CTAs had 36 true aneurysms. Ten of 36 (27.8%) true aneurysms flagged by Viz ANEURYSM were missed by physicians interpreting the CTAs. This study's implications underscore a recurring issue in routine clinical care, as UCAs that necessitate follow‐up are frequently overlooked. The implementation of a machine learning algorithm to flag potential aneurysms in imaging studies presents a promising solution to capture those cases that might otherwise be neglected. The study's sample encompasses UCAs from a diverse spectrum of clinical setting, ranging from comprehensive stroke centers to free stand‐alone emergency rooms. However, the analysis does not shed light on whether patients who received referrals for follow‐up differed in any way from those who did not. Factors such as short life expectancy, underlying comorbidities, or compliance challenges may have contributed to the variation in follow‐up rates but remain unexplored in this analysis.

The study also does not report the software's false‐negative rate, leaving us without insight into how many aneurysms the machine learning algorithm missed but were correctly identified by radiologist physicians during their study assessments. This is a significant limitation that hampers our understanding of the specificity of Viz ANEURYSM. In the subgroup of 50 UCAs examined, a substantial 38% (19/50) of CTAs were erroneously flagged as UCAs by the software. This included cases such as arteriovenous malformations, infundibula, and even 14 normal studies. This misclassification underscores the need for further refinement of the algorithm, particularly given the intricate variations of the circle of Willis, which continue to present challenges even for experienced human readers.

Furthermore, the study leans on the practical risk score (PHASES) as a benchmark for predicting the risk of rupture. However, it is important to approach this aspect of the study with caution, as several reports have raised questions about the reliability of the PHASES score as a ground truth in this context.⁷

The demand for improved tools for detecting UCAs is of paramount importance, especially considering the substantial volume of cerebral angiographic studies performed across various hospital settings. Furthermore, the shortage of qualified neuroradiologists has compelled the adoption of remote third‐party services,⁸ limiting direct communication with the referring provider, and therefore foreshortening the exchange of meaningful clinical information that may guide the diagnosis. We commend the authors for their diligent effort in studying the potential of machine learning in detecting UCAs. The integration of artificial intelligence into routine patient care is already under way, and its future prospects are promising. This technological advancement is poised to enhance health care outcomes significantly.