Articles In Press
Back to the articles list |
Back to browse issues page
Ethics code: IR.IAU.BIRJAND.REC.1404.036
Department of Computer Engineering, Fe.C., Islamic Azad University, Ferdows, Iran
Abstract: (20 Views)
Objective: Chest radiography is a vital tool for disease diagnosis; however, it is susceptible to human error due to the subtlety of findings and their overlap with complex anatomical structures. The objective of this research is to present a new framework, YOLO-CA-NET, to improve the detection accuracy and localization of subtle and small abnormalities in radiographic images using advanced attention mechanisms.
Methods: In this study, the YOLOv11 model was upgraded by strategically integrating the Coordinate Attention module into its backbone blocks to extract position-sensitive information and preserve the spatial details of small lesions. For training and evaluation, the standard VinDr-CXR dataset, comprising 15,000 radiographic images with annotations for 14 common types of abnormalities, was utilized. Pre-processing included normalization and the CLAHE technique to enhance the contrast of low-visibility findings.
Results: The research findings indicated that the proposed YOLO-CA-NET model achieved a mean Average Precision (mAP) of 0.329 and a mean Recall of 0.514. These results demonstrate a 9.3% increase in precision and a 19% increase in recall compared to the baseline YOLOv11 model. This improvement validates the high potential of the proposed framework in detecting and localizing low-contrast and subtle lesions that might be overlooked by the baseline model.
Conclusion: Integrating the Coordinate Attention mechanism into the YOLOv11 architecture significantly enhances the model's ability to understand long-range spatial dependencies and precisely localize subtle abnormalities. The proposed framework can serve as an efficient computer-aided diagnosis tool to mitigate human errors arising from missing small lesions in clinical settings.
Methods: In this study, the YOLOv11 model was upgraded by strategically integrating the Coordinate Attention module into its backbone blocks to extract position-sensitive information and preserve the spatial details of small lesions. For training and evaluation, the standard VinDr-CXR dataset, comprising 15,000 radiographic images with annotations for 14 common types of abnormalities, was utilized. Pre-processing included normalization and the CLAHE technique to enhance the contrast of low-visibility findings.
Results: The research findings indicated that the proposed YOLO-CA-NET model achieved a mean Average Precision (mAP) of 0.329 and a mean Recall of 0.514. These results demonstrate a 9.3% increase in precision and a 19% increase in recall compared to the baseline YOLOv11 model. This improvement validates the high potential of the proposed framework in detecting and localizing low-contrast and subtle lesions that might be overlooked by the baseline model.
Conclusion: Integrating the Coordinate Attention mechanism into the YOLOv11 architecture significantly enhances the model's ability to understand long-range spatial dependencies and precisely localize subtle abnormalities. The proposed framework can serve as an efficient computer-aided diagnosis tool to mitigate human errors arising from missing small lesions in clinical settings.
Keywords: Thoracic Radiography, Subtle Pathologies, YOLOv11, Coordinate Attention Mechanism, Object Detection.
Send email to the article author
| Rights and permissions | |
 |
This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. |
Contact Us
Post Address: Bandar Abbas, Imam Hossein Boulevard, campus of University of Medical Sciences, School of Nursing and Midwifery.
Tel: +(98) 763 331 7723
Cell Phone: +(98) 939 654 9816
jmis
hums.ac.ir
