TFL-Net: A Hybrid Deep Learning Framework for Tibia Fracture Detection and Localization

Abstract

Tibia fractures present significant diagnostic challenges in orthopedics due to their subtle presentation and variable healing patterns, often resulting in delayed treatment and adverse outcomes for athletes and patients. This work presents a comprehensive, hybrid deep learning framework called TFL-net (Tibia Fracture detection and Localization network) for tibia fracture diagnosis, with a particular emphasis on rapid and precise localization using the YOLOv8 object detection model. TFL-net integrates three key methodologies: (1) VGG16-based deep feature extraction combined with a Support Vector Machine (SVM) classifier for accurate fracture classification; (2) a custom-trained YOLOv8 network for real-time localization of fracture regions, leveraging advanced spatial attention, multi-scale feature integration, and grid-based prediction to achieve a mean average precision (mAP@0.5) of 0.964 and an inference time of 21.6 ms per image; and (3) an Extreme Gradient Boosting (XGBoost) module for individualized healing time prediction based on patient-specific clinical and demographic parameters. The YOLOv8 model, central to the localization task, demonstrates robust detection performance and efficiency, making it suitable for clinical deployment. The integrated system, accessible via a graphical user interface, streamlines the workflow from initial diagnosis to prognosis, supporting clinicians with accurate, real-time insights and actionable treatment planning for tibia fractures.

Ähnliche Arbeiten