Wrist fracture detection using self-supervised learning methodology

Abstract

Objectives: This study aimed to assist radiologists in faster and more accurate diagnosis by automating bone fracture detection in pediatric trauma wrist radiographic images using self-supervised learning. This addresses data labeling challenges associated with traditional deep learning models in medical imaging. Methods: In this study, we trained the model backbone for feature extraction. Then, we used this backbone to train a complete classification model for classifying images as fracture or non-fracture on the publically available Kaggle and GRAZPERDWRI-DX dataset using ResNet-18 in pediatric wrist radiographs. Results: The resulting output revealed that the model was able to detect fracture and non-fracture images with 94.10% accuracy, 93.21% specificity, and an area under the receiver operating characteristics of 94.12%. Conclusion: This self-supervised model showed a promising approach and paved the way for efficient and accurate fracture detection, ultimately enhancing radiological diagnosis without relying on extensive labeled data.