Benchmarking YOLO Models for Intracranial Hemorrhage Detection Using Varied CT Data Sources

• Published in: IEEE access 2024, Vol.12, p.188084-188101
• Main Authors: Tapia, Gonzalo, Allende-Cid, Hector, Chabert, Steren, Mery, Domingo, Salas, Rodrigo
• Format: Article
• Language: English
• Subjects: Accuracy; Computed assisted diagnosis (CAD); Computed tomography; Convolutional neural networks; convolutional neural networks (CNN); Data models; Hemorrhaging; intracranial hemorrhage detection; medical imaging; Microprocessors; Predictive models; Solid modeling; Training; YOLO
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2024.3510517

Intracranial hemorrhages (ICH) are a significant challenge in emergency medicine due to the critical nature of a timely and accurate diagnosis. This study evaluates the performance of six versions of the You Only Look Once (YOLO) object detection model, from YOLOv5 to YOLOv10, in detecting ICH using computed tomography (CT) scans. The primary focus is understanding the advancements in YOLO architectures over time and their impact on detection accuracy and inference speed. The study used the Brain Hemorrhage Extended Dataset (BHX), comprising 491 CT scans with annotations for six types of hemorrhages: epidural, subdural, subarachnoid, intraparenchymal, intraventricular, and chronic hemorrhage, and introduces a new data set obtained from a major hospital in Chile. The models were trained using a combination of single-class and multi-class approaches to address class imbalance and were evaluated based on precision, recall, F1 score, and mean average precision (mAP). The models were evaluated in three distinct contexts: 1) a biased scenario where images of the same individual could appear in both training and testing sets, 2) a cross-validation setup ensuring the independence of images by separating the sets based on subjects, and 3) an external validation using one dataset for training and the Chilean dataset for testing, maintaining full independence between training and evaluation. The findings indicate that YOLOv8 and YOLOv10 demonstrate superior detection accuracy and inference efficiency performance, respectively, compared to previous versions. In particular, with image independence, YOLOv8 reached the highest average mAP for all classes, with a score of 0.4. This comparative analysis provides information on the effectiveness of architectural advances in YOLO models for medical applications and suggests directions for future improvements in ICH detection.