Hierarchical Bone Fracture Detection and Classification Using YOLOv9 with Grad-CAM++ Visualization

This paper describes the Bone Fracture Detection System using the YOLOv9 model. It accurately classifies and detects fractures in different bones, including the femur, humerus, radius, ulna, tibia, fibula, elbow, fingers, forearm, shoulder, and wrist. A hierarchical approach was designed so that the class of bone, anatomy, fracture detection, and classification would come in sequence. The system was trained and validated on the Human Bone Fractures Multi-modal Image Dataset (HBFMID) supplemented with additional X-ray images for better representation of bone type and fracture pattern. Advanced preprocessing, including image augmentation and normalization, improved generalization. This YOLOv9 architecture is optimized for advanced feature extraction and localization. The system is tested using performance metrics such as the mean Average Precision (mAP) and Intersection over Union (IoU). Classification of bone type was set at 96.50% mAP@0.5 and fracture detection as 93.50% mAP@0.5. These results indicate how the system changes diagnostics by facilitating fast, reliable, and automatic fracture detection. This helps eliminate some errors and accelerates treatment planning, pointing out the need for artificial intelligence in healthcare.