Improved Bone Fracture Classification Using ConvNextV2 with Optimized GRN and Meta-Ensemble

Abstract

This study introduces a hybrid framework for classifying bone fractures from X-ray images, aiming to enhance diagnostic accuracy and interpretability. Utilizing the Bone Fracture Multi-Region X-ray Dataset, we fine-tune the ConvNeXt V2 model for binary classification. Features extracted from the CNN are optimized using RFECV and employed to train an ensemble of classifiers, with a Random Forest meta-learner integrating their predictions. Our method achieves a test accuracy of 99.21%, with precision, recall, and F1-score of 98.53%, 100%, and 99.26%, respectively. Interpretability is addressed through Grad-CAM and saliency maps, providing insights into the model's decision-making process. This approach marks a significant step forward in automated fracture diagnosis, combining high performance with transparency essential for clinical trust and adoption.

Ähnliche Arbeiten