A Lightweight One-Dimensional CNN Framework For Intracranial Aneurysm Detection Using Derived Clinical and Morphological Features From CTA and MRA

Abstract

Intracranial aneurysms (IAs) are outpouchings of the cerebral arteries that can rupture at any time, causing an often fatal and debilitating neurological emergency called subarachnoid hemorrhage. Computed Tomography Angiography (CTA) and Magnetic Resonance Angiography (MRA) provides the early diagnosis that significantly ameliorates clinical outcomes, but manual interpretation is labour intensive, subject to observer variability, and increasingly difficult as the number of neuroimaging examinations increases. Designed as a novel multi-modal neuroimaging based deep learning approach, we present here an automated computer-aided detection (CAD) framework for binary intracranial aneurysm detection from brain MRI scans. It introduces a unified preprocessing pipeline that can normalize intensities and anatomy as well as harmonize data from both Digital Imaging and Communications in Medicine (DICOM) and Neuroimaging Informatics Technology Initiative (NIfTI) image formats. Methods: We trained a one-dimensional batch normalization, dropout regularization and adaptive feature normalization CNN for aneurysm classification. Receiver operating characteristic (ROC) analysis, precision-recall analysis, confusion matrix evaluation and anatomical site-wise performance assessment was done to evaluate model performance. Our framework had an AUROC of 0.667 and AUPRC-0.556 while demonstrating moderate discriminative capability yet outperformed random classification. The confusion matrix analysis revealed 1,760 true negatives, 990 true positives,725 false positives and 873 false negatives resulting in sensitivity of ~53.1%, specificity of 70.8%, precision of 57.7% and balanced accuracy of 61.9%. Analysis of the anatomica site showed variation in detection performance across cerebrovascular regions. We incurred to 4.02–2 in obtaining no quantitative assessment of subgroup performance by age, sex and imaging modality i.e. for future validation the current dataset identifies an important direction towards analysis. While diagnostic performance remains below clinical-grade levels with this framework, results indicate the feasibility of deep learning classification with standardized multi-modal preprocessing for automated screening of intracranial aneurysms. The solution can serve as a clinical decision-support system to help radiologists by prioritizing suspicious examinations and lowering interpretation burden. The embedding is a feature-based representation that dramatically decreases computational complexity while still retaining clinically relevant information for classification of aneurysms. Future work will address three-dimensional feature learning, lesion localization, explainable artificial intelligence, multicenter external validation and advanced transformer-based architectures to enhance their diagnostic performance for clinical use.