YOLOv11-based multi-task learning for enhanced bone fracture detection and classification in X-ray images

Abstract

Objective

This study presents a multi-task learning framework based on the YOLOv11 architecture to improve both fracture detection and localization. The goal is to provide an efficient solution for clinical applications.

Materials and methods

We used a large dataset of X-ray images, including both fracture and non-fracture cases from the upper and lower extremities. The dataset was divided into three parts: training (70%), validation (15%), and test (15%). The training set had 10,966 cases (5778 normal, 5188 with fractures), while the validation and test sets each contained 2350 cases (1238 normal, 1112 with fractures). A multi-task learning model based on YOLOv11 was trained for fracture classification and localization. We applied data augmentation to prevent overfitting and improve generalization. Model performance was evaluated using two metrics: mean Average Precision (mAP) and Intersection over Union (IoU), with comparisons made to Faster R-CNN and SSD models. Training was done with a learning rate of 0.001 and a batch size of 16, using the Adam optimizer for better convergence. We also benchmarked the YOLOv11 model against Faster R-CNN and SSD to assess performance using mAP and IoU scores at different thresholds.

Results

The YOLOv11 model achieved excellent results, with a mean Average Precision (mAP) of 96.8% at an IoU threshold of 0.5 and an IoU of 92.5%. These results were better than Faster R-CNN (mAP: 87.5%, IoU: 85.23%) and SSD (mAP: 82.9%, IoU: 80.12%), showing that YOLOv11 outperformed these models in fracture detection and localization. This improvement highlights the model's strength and efficiency for real-time use.

Conclusions

The YOLOv11-based multi-task learning framework significantly outperforms traditional methods, offering high accuracy and real-time fracture localization. This model shows great potential for clinical use, improving diagnostic accuracy, increasing productivity, and streamlining the workflow for radiologists.