nnU-Net Based Architecture for the Detection and Localization of Brain Aneurysm

Abstract

A brain aneurysm is an enlargement in one of the blood vessels within the brain, although it may rupture and bring fatality. On average, approximately $76,000-200,000$ cases of cerebral aneurysm are believed to occur each year in India, and most victims of ruptures die, with many survivors permanently disabled by the incident, making early diagnosis vital. In this paper, we introduce an entire aneurysm early detection and localization medical image pipeline. A preprocessing and filtering of input data is done to guarantee consistency, and a coarse-to-fine method of vessel separation is employed to extract the significant vascular areas. This assists the model to only learn significant structures of the anatomy. Lastly, it predicts the presence of aneurysms and location using vessel-guided features in a 3D classification model, which enhances the results of aneurysm detection because a slice-toslice spatial context is utilized. The accuracy, precision, recall, and F1- score of the proposed method are 91.30, 92.1, 88.30 and 90.20, respectively, with an AUC-ROC of 0.911, which indicates a high overall performance. On localization, the model gives an accuracy of 0.960 and a recall rate of 0.738, which means that the model can detect the aneurysm regions and cover well. These findings expose that the developed pipeline presents high reliability and can be used in early detection to help clinicians make better decisions and improve patient outcomes.