Automated diagnosis and classification of metacarpal and phalangeal fractures using a convolutional neural network: a retrospective data analysis study.

Abstract

Background and purpose: Hand fractures are commonly presented in emergency departments, yet diagnostic errors persist, leading to potential complications. The use of artificial intelligence (AI) in fracture detection has shown promise, but research focusing on hand metacarpal and phalangeal fractures remains limited. We aimed to train and evaluate a convolutional neural network (CNN) model to diagnose metacarpal and phalangeal fractures using plain radiographs according to the AO/OTA classification system and custom classifiers.

Methods: A retrospective analysis of 7,515 examinations comprising 27,965 images was conducted, with datasets divided into training, validation, and test datasets. A CNN architecture was based on ResNet and implemented using PyTorch, with the integration of data augmentation techniques.

Results: The CNN model achieved a mean weighted AUC of 0.84 for hand fractures, with 86% sensitivity and 76% specificity. The model performed best in diagnosing transverse metacarpal fractures, AUC = 0.91, 100% sensitivity, 87% specificity, and tuft phalangeal fractures, AUC = 0.97, 100% sensitivity, 96% specificity. Performance was lower for complex patterns like oblique phalangeal fractures, AUC = 0.76.

Conclusion: Our study demonstrated that a CNN model can effectively diagnose and classify metacarpal and phalangeal fractures using plain radiographs, achieving a mean weighted AUC of 0.84. 7 categories were deemed as acceptable, 9 categories as excellent, and 3 categories as outstanding. Our findings indicate that a CNN model may be used in the classification of hand fractures.