Enhanced Bone Fracture Prediction Using Transfer Learning and YOLOv8 Deep Learning Models

Bone fractures are among the most common injuries encountered in clinical practice, yet their accurate identification in medical imagery remains challenging due to variations in image quality, bone morphology, and fracture subtlety. This work presents a deep learning–based approach for automated bone fracture detection using three enhanced YOLO architectures. The proposed system, BoneTrack AI, leverages transfer learning from large-scale object detection datasets to improve generalization on medical images. Three models are developed and compared: a standard YOLOv8 baseline, a YOLOv12s variant augmented with a Global Attention Module that sequentially recalibrates channel and spatial features, and a YOLOv26n model incorporating a learnable multi‑scale fusion block to adaptively combine pyramidal feature representations. All models are trained on a curated dataset containing two categories—fracture and non‑fracture—enabling binary classification and localization. Extensive evaluation demonstrates the strengths of attention and fusion mechanisms for focusing on diagnostically relevant regions. The best‑performing model is deployed as the backend of a web‑based application, allowing users to upload medical images and receive immediate fracture predictions with visual overlays. This integrated solution illustrates how modern object detection architectures, combined with targeted architectural enhancements, can assist clinicians by providing rapid, interpretable second opinions.

Keywords: Bone fracture detection, YOLOv8, transfer learning, attention mechanism, multi‑scale feature fusion, medical image analysis