# Worthwhile or Not? The Pain–Gain Ratio of Screening Routine cMRIs in a Maximum Care University Hospital for Incidental Intracranial Aneurysms Using Artificial Intelligence

## TL;DR

AI screening of routine brain MRIs can detect previously missed aneurysms but may lead to more follow-up tests for minor findings.

## Contribution

This study evaluates the clinical value and cost of AI-based incidental aneurysm detection in routine cMRIs using a large hospital cohort.

## Key findings

- AI detected aneurysms with high sensitivity, reducing non-reporting by radiologists.
- Routine AI screening would require 22 cMRIs to prompt further imaging and 221 to detect a clinically relevant aneurysm.
- AI alerts led to additional imaging for 4–5% of patients, with 0.45–0.74% found to have previously undetected aneurysms.

## Abstract

Background: Aneurysm-related subarachnoid hemorrhage is a life-threatening form of stroke. While medical image acquisition for aneurysm screening is limited to high-risk patients, advances in artificial intelligence (AI)-based image analysis suggest that AI-driven routine screening of imaging studies acquired for other clinical reasons could be valuable. Methods: A representative cohort of 1761 routine cranial magnetic resonance imaging scans [cMRIs] (with time-of-flight angiographies) from patients without previously known intracranial aneurysms was established by combining 854 general radiology 1.5T and 907 neuroradiology 3.0T cMRIs. TOF-MRAs were analyzed with a commercial AI algorithm for aneurysm detection. Neuroradiology consultants re-assessed cMRIs with AI results, providing Likert-based confidence scores (0–3) and work-up recommendations for suspicious findings. Original cMRI reports from more than 90 radiologists and neuroradiologists were reviewed, and patients with new findings were contacted for consultations including follow-up imaging (cMRI / catheter angiography [DSA]). Statistical analysis was conducted based on descriptive statistics, common diagnostic metrics, and the number needed to screen (NNS), defined as the number of cMRIs that must be analyzed with AI to achieve specific clinical endpoints. Results: Initial cMRI reporting by radiologists/neuroradiologists demonstrated a high risk of incidental aneurysm non-reporting (94.4% / 86.4%). A finding-based analysis revealed high AI algorithm sensitivities (100% [3T] / 94.1% [1.5T] for certain aneurysms of any size, well above 90% for any suspicious findings > 2 mm), associated with AI alerts triggered in 22% of cMRIs with PPVs of 7.5–25.2% (depending on the inclusion of inconclusive findings). The NNS to prompt further imaging work-/follow-up was 22, while the NNS to detect an aneurysm with a possible therapeutic impact was 221. Reference readings and patient consultations suggest that routine AI-driven cMRI screening would lead to additional imaging for 4–5% of patients, with 0.45% to 0.74% found to have previously undetected aneurysms with possibly therapeutic implications. Conclusions: AI-based second-reader screening substantially reduces incidental aneurysm non-reporting but may disproportionally increase follow-/work-up imaging demands also for minor or inconclusive findings with associated patient concern. Future research should focus on (subgroup-specific) AI optimization and cost-effectiveness analyses.