Comprehensive Summary
This study introduces a hybrid diagnostic framework that combines radiomics and deep learning to enhance the detection and classification of wrist fractures on X-ray images. Utilizing a dataset of 3,537 X-rays, the authors applied radiomic feature extraction via PyRadiomics and deep feature learning through an attention-enhanced autoencoder. The hybrid feature set underwent preprocessing, normalization, and dimensionality reduction using techniques like Mutual Information, PCA, ANOVA, and Recursive Feature Elimination. Classification was performed using ensemble models, with the Voting Classifier achieving the best performance at 95% accuracy, 94% sensitivity, and 96% AUC-ROC. The system's robustness was validated through external testing across three independent centers, with accuracies ranging from 90.8% to 92.6% and AUC-ROC between 92.7% and 94.2%. Attention maps and SHAP values confirmed model interpretability, highlighting clinically relevant features.
Outcomes and Implications
The research presents a clinically viable AI framework that addresses limitations in current diagnostic tools by integrating reproducible radiomic features with robust, attention-guided deep learning. The hybrid model significantly outperforms traditional AI systems in accuracy and generalizability, achieving 95% accuracy and 96% AUC. This is crucial for minimizing missed diagnoses. The ensemble approach, especially the Voting Classifier with Mutual Information, supports accurate fracture classification across varied image quality and clinical settings. The model's high cross-site stability and low false-positive/false-negative rates enhance its reliability. Its modular design and generalizability position it for expansion into other fracture types or anatomical regions, making it a scalable and interpretable AI solution for musculoskeletal radiology.