Browsing by Author "CIIT/SP24-RCS-020/LHR"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • No Thumbnail Available
    Item
    Advancing Pediatric Wrist Fracture Detection with YOLOv12: Exploiting Attention Mechanisms and Transfer Learning
    (Library Information Services, COMSATS University Islamabad, Lahore Campus, 2025) Maham Nasir; CIIT/SP24-RCS-020/LHR; Dr. Aksam Iftikhar, 69 P; LHR TP 10041
    Pediatric wrist fractures are among the common musculoskeletal injuries that have been experienced in emergency and radiology departments. X-ray images diagnosis would be crucial to preventing long-term effects of malunion, chronic pain, and functional impairment. Nevertheless, the pediatric X-rays of the wrist are not able to easily identify fractures because of fine fracture fissures, growth plates, and structural fusion, as well as, the inconsistency in image quality. Even though methods based on deep learning have shown good performance in fracture detection, most of the available methods are confined to classification tasks or fail to perform sufficiently sensitive and accurate localization on small childhood fractures. In this thesis, an improved YOLOv12-based attention-enhanced model is proposed to detect and localize pediatric wrist fractures in X-ray images in an automated manner. The suggested strategy takes advantage of the newly introduced YOLOv12 with its Area Attention (A2) modules, Residual Efficient Layer Aggregation Networks (R-ELAN), and transfer learning that enhance the discrimination of features and multi-scale representation. To ensure the quality of X-ray images and make the models more resistant to exposure, contrast, and anatomy changes, a complex medical-specific preprocessing and data-augmentation pipeline is used. The model is trained and assessed with the GRAZPEDWRI-DX publicly accessible dataset that comprises of more than 20,000 pediatric wrist X-ray images with accurate annotations. Massive experimental tests are performed based on various input resolutions and variants of YOLO, and the performance is measured in terms of precision, recall, mean Average Precision (mAP50 and mAP50-95), and the confusion matrix. The proposed model performs much better than state-of-the-art YOLO-based methods with a precision of 0.797 and a recall of 0.565, mAP at 50 of 0.669, and mAP at 50-95 of 0.381