Comparative analysis of ResNet, ResNet-SE, and attention-based RaNet for hemorrhage classification in CT images using deep learning

Intracranial hemorrhage (ICH) is a critical medical condition associated with blood vessel rupture, demanding prompt intervention for optimal outcomes. This study focuses on evaluating the classification performance of hemorrhage detection and grading architectures based on Residual Networks (HResNet) in the context of computed tomography (CT) scans. Leveraging a comprehensive dataset of 22,811 images sourced from 491 scans within the CQ500 dataset, this research investigates the effectiveness of HResNet, ResNet with squeeze and-excitation (HResNet-SE), and ResNet with attention (HRaNet) architectures. The results of this study demonstrate that HRaNet stands out as the superior performer, showcasing remarkable metrics across various evaluation criteria. Notably, HRaNet achieves an impressive Area Under the Curve (AUC) exceeding 0.96, a Jaccard index of 0.9130, a Macro-F1 score of 0.9464, and a Micro-F1 score of 0.9545 for subtype grading. Moreover, HRaNet maintains its robust reliability, as evidenced by a Kappa score surpassing 0.9063. The findings underscore the potential of deep learning models to assist radiologists in the accurate diagnosis of ICH, thus enhancing patient care. Additionally, the study's consideration of prediction time and parameter count provides practical insights into selecting optimal models for clinical integration. Given the complex nature of hemorrhages, which can occur in various locations this study addresses the multi-label classification challenge inherent in detecting multiple pathologies from a single CT scan. The study contributes to the growing body of knowledge in medical image analysis, facilitating advancements in early diagnosis and management of critical medical conditions.