E5108. Role of Artificial Intelligence in Detection of Intracranial Aneurysms: Clinical Applications
  1. Delaram Shakoor; Yale
  2. Ajay Malhotra-; ; Yale
Intracranial aneurysms (IAs) are a major health concern. An efficient method for early detection of IAs is required to provide appropriate radiological screening guidelines. Application of artificial intelligence (AI) for detection of neuroradiology emergencies, such as aneurysms, has received recent attention. Several studies have reported the high sensitivity of AI for detection of small aneurysms; however, data regarding its specificity are scarce. Besides, the rupture prediction model and the clinical impact of identifying AI-detected small aneurysms are unknown. Therefore, we intend to evaluate the size of AI-detected aneurysms that were not mentioned in the radiology report and assess the interreader agreement of second reads by different readers.

Materials and Methods:
We reviewed 2100 consecutive CTAs performed in our center, and the studies that were flagged positive by the AI software (AiDoc) were extracted. The reports of these studies were reviewed to identify those studies in which the radiology report did not mention the aneurysms. The identified 30 studies were read anonymously by three neuroradiology attendings and a neuroradiology fellow. The average of the size of each aneurysm was calculated, and interreader agreements were calculated using the intraclass correlation (ICC).

Reader 1 identified 21 (70%) aneurysms, with the mean and range of 2.4 ± 0.7 (1.5–4) mm. Reader 2 detected 14 (47%) aneurysms, with the mean and range of 2.5 ± 1.3 (1–5) mm. Reader 3 found 6 (20%) aneurysms with the mean and range of 2.3 ± 0.8 (2–4) mm, and Reader 4 noted 15 (50%) aneurysms with the mean and range of 1.7 ± 1 (1.5–5) mm. The ICC between readers was 55.2% (95% CI: 25–76%, p = 0.001). In six (20%) cases, none of the readers identified any aneurysm, while in five (17%) cases, three readers did not identify any aneurysm. In four (13%) cases, no reader identified the same area concerning for aneurysms.

Further attempts should be made to improve diagnostic accuracy of AI for detection of aneurysm, with emphasis on specificity. The focus of future AI research should be made on the improving prediction models for aneurysm rupture, which will be of more clinical significance.