MoE-Net: A deep ensemble framework optimized by genetic algorithm for nasal bone fracture detection on lateral X-ray images

Reliable identification of nasal bone fractures on lateral radiographs is a persistent clinical challenge, largely due to the subtle radiographic presentation of fractures, variability and noise inherent in routine imaging, and the limited availability of expert-annotated datasets. To overcome these obstacles, we present MoE-Net, a precision-oriented Mixture-of-Experts Network designed to enhance automated nasal bone fracture detection. The proposed approach systematically evaluates a pool of ten pre-trained convolutional neural network (CNN) and transformer-based architectures, from which the three most accurate and error-resilient models, InceptionResNetV2, DenseNet121, and Xception, are selected as specialized experts. Each expert model undergoes targeted fine-tuning, while a genetic algorithm optimizes their contribution weights within the ensemble to maximize predictive performance. MoE-Net demonstrates substantial performance gains over individual models and existing handcrafted-feature-based methods. On the test dataset, the framework achieves an accuracy of 91.70%, sensitivity of 91.70%, specificity of 89.95%, F2-score of 91.69%, Matthews Correlation Coefficient of 83.48%, and an area under the ROC curve of 91.99%. This performance reflects a clinically meaningful balance between minimizing false negatives, critical for preventing missed diagnoses, and controlling false positives to avoid unnecessary interventions. These findings support the clinical applicability of MoE-Net as a robust, high-performance decision-support tool for nasal bone fracture detection. The study highlights the advantages of precision-driven ensemble strategies in medical imaging and underscores their potential to improve diagnostic accuracy and contribute to more efficient patient care. • A precision-driven Mixture-of-Experts framework (MoE-Net) is proposed for nasal bone fracture detection from lateral X-ray images. • Three complementary CNN experts are selected and fine-tuned to minimize clinically critical diagnostic errors. • A genetic algorithm optimizes expert contribution weights, improving ensemble reliability and diagnostic balance. • MoE-Net achieves high performance (91.9% accuracy, 91.9% sensitivity, MCC 83.9%) on expert-annotated radiographs. • The framework reduces missed fractures while controlling false positives, supporting real-time clinical deployment.