Clasificación de fracturas en extremidades superiores con aprendizaje profundo

Accurate identification of upper extremity fractures is essential for timely and reliable diagnosis in emergency medical settings. This study evaluates and compares the performance of three pre-trained deep learning architectures: EfficientNet- B4, ResNet-50, and ConvNeXt-Large, applied to the automatic classification of bone fractures in radiographic images from the MURA repository, encompassing seven anatomical regions. Advanced image preprocessing techniques, including Unsharp Masking and Contrast-Limited Adaptive Histogram Equalization (CLAHE), were employed in conjunction with data normalization and balancing strategies. The models were trained in two experimental setups: a binary classification distinguishing between “fracture” and “non-fracture” images, and a multiclass configuration identifying 14 distinct fracture types. Performance evaluation using F1-Score, sensitivity, accuracy, and ROC–AUC metrics demonstrated that ConvNeXt-Large achieved the highest overall results, reaching accuracies of 99.0% in binary classification and 99.4% in multiclass classification. These findings position ConvNeXt-Large as a highly promising tool for supporting early and precise fracture diagnosis.