Leveraging convolutional neural networks to address overfitting and generalizability in automated bone fracture detection

Bone fractures represent a significant health burden that demands precise and timely diagnosis to optimize patient outcomes. To address challenges such as data scarcity, overfitting, and generalizability, this study investigates the use of convolutional neural networks (CNNs) for automated fracture detection in X-ray images. A dataset of 4,900 X-ray images was preprocessed and evenly divided into training, validation, and test subsets. The proposed CNN model directly addressed generalizability and overfitting issues by prioritizing training stability and incorporating advanced techniques. These techniques included batch normalization and dropout to enhance stability and mitigate overfitting, with five-fold cross-validation yielding an average accuracy of 95%. Validation and held-out test datasets achieved accuracies of 95.8% and 94.5%, respectively, while external validation on an independent dataset confirmed the model’s generalizability at 91.7%. High recall rates across all datasets underscore the model’s capacity to minimize missed fracture diagnoses, whereas slightly lower precision on external data indicates a need to address false positives. These findings suggest that artificial intelligence is best deployed as a screening tool, serving as an initial triage mechanism that flags potential cases for further human-guided evaluation, thereby enhancing clinical efficiency without replacing the diagnostic expertise of healthcare professionals.