184 Development of an Autonomous Artificial Intelligence Tool for the Detection and Morphological Characterization of Unruptured Intracranial Aneurysms

INTRODUCTION: Several artificial intelligence (AI) models for detecting unruptured intracranial aneurysms (UIAs) have been developed in recent years, however, all have performed poorly on external validation, and none are able to obtain key morphological parameters. These parameters have consistently outperformed the traditional clinical variables in predicting rupture. METHODS: We collected CTA scans of UIA at our institution and at one international institution. Inclusion criteria included scans with untreated UIA without subarachnoid hemorrhage, slice thickness < 2 mm, and acceptable quality. The scans were all manually segmented on 3DSlicer and augmented to amplify the dataset size. We used these scans to develop and train a novel semantic segmentation deep learning model which was then internally and externally vidated. RESULTS: Between September 2000 and January 2024, there were 1940 CTA scans meeting the inclusion criteria with 2450 aneurysms. Most aneurysms (90.4%) were in the anterior circulation with the average maximum diameter being 5.1 mm. These scans were successfully used to create and train the novel algorithm. The model was able to achieve an intersection-over-union (IOU) of 95% and accuracy of 99% on internal validation. On external validation, the model achieved an IOU of 91% and accuracy of 93%. In terms of measurement, the model had 98% volume concordance with the ground truth. Clinically, the results of the model were deemed excellent by senior neurosurgeons. CONCLUSIONS: We present the first AI-based model which autonomously detects UIAs and retrieves key morphological parameters on CTA scans. This model now allows cerebrovascular specialists to integrate these predictors with ease into their decision-making.