Optimizing YOLOv9 for automated detection of stroke lesions in brain CT images

• Published in: Neural computing & applications Vol. 37; no. 29; pp. 24169 - 24189
• Main Authors: Talaat, Fatma M., Shaban, Warda M.
• Format: Journal Article
• Language: English
• Published: London Springer London 01.10.2025; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Annotations; Artificial Intelligence; Automation; Brain; Brain cancer; Brain research; Computational Biology/Bioinformatics; Computational Science and Engineering; Computed tomography; Computer Science; Computer vision; Data Mining and Knowledge Discovery; Datasets; Deep learning; Diagnosis; Explainable artificial intelligence; Image Processing and Computer Vision; Ischemia; Lesions; Literature reviews; Machine learning; Medical imaging; Methods; Neural networks; Object recognition; Optimization; Original Article; Patients; Physicians; Probability and Statistics in Computer Science; Real time; Stroke; Support vector machines; YOLOv9; Brain CT imaging; Explainable AI (XAI); Stroke detection
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-025-11560-4

The prompt detection and accurate localization of stroke-induced cerebral damage in Computed Tomography (CT) scans are essential for optimal treatment and patient prognosis. Traditional techniques significantly depend on the proficiency of radiologists, which can be labor-intensive and susceptible to inaccuracies. This study presents a novel methodology utilizing the YOLOv9 deep learning architecture, termed StrokeYOLO, for the automated detection of stroke lesions in brain CT images. Although YOLOv9 is primarily employed for general object detection, we have modified it to specifically focus on the nuanced characteristics of stroke lesions. Through the modification of anchor boxes, the refinement of feature extraction, and the fine-tuning of the model, we have improved its capacity to accurately identify smaller stroke regions. The proposed model was trained and assessed on a dataset of annotated brain CT scans, exhibiting outstanding efficacy in identifying ischemic and hemorrhagic stroke regions. The detection accuracy was further corroborated against expert annotations, attaining results that were comparable to or superior to traditional methods. Furthermore, we utilized explainable AI methodologies to enhance transparency in the model's decision-making process, thereby promoting trust and clinical implementation. This study emphasizes the capabilities of YOLOv9 as a real-time, automated instrument for facilitating stroke diagnosis while tackling the challenges and prospects of utilizing object detection models in medical imaging. StrokeYOLO attained an accuracy of 98.7%, precision of 99.92%, recall of 99.1%, and F 1-measure of 99.51% on the evaluation dataset, surpassing current methodologies. The findings underscore the capability of YOLOv9 as a real-time, automated instrument for facilitating stroke diagnosis while tackling the challenges and prospects of utilizing object detection models in medical imaging.