A hybrid deep learning approach for detection and classification of pediatric wrist pathologies in X-ray images

Pediatric wrist injuries are common cases encountered in hospital emergency departments, with fractures being the most prevalent among them. Traditionally, radiologists and surgeons collaborate in diagnosing such injuries. However, recent advancements in deep learning-based algorithms have shown promise in automating and expediting the diagnostic process. This thesis proposes a hybrid method for the classification and localization of various pediatric wrist injuries, including bone anomaly, bone lesion, foreign body, fracture, metal, periosteal reaction, pronator sign, and soft tissue injuries. Using the GRAZPEDWRI-DX dataset, the proposed method integrates the YOLOv8 algorithm for object detection with ensemble-based convolutional neural networks and the attention mechanism-based Vision Transformer (ViT) for classification. For the localization of pediatric wrist injuries, the YOLOv8x model achieved results of 83.4% precision, 77.8% recall, 84.7% mAP50, and 65.7% mAP50-95. In the classification phase, the hybrid use of Xception, DenseNet201, and EfficientNetB5 models in an ensemble framework combined with a Vision Transformer yielded a remarkable accuracy of 96.29%. The proposed method and the results obtained in this thesis have the potential to guide future research toward faster and more accurate diagnosis of pediatric wrist injuries.