Automated detection of pediatric forearm fractures in X-ray images using deep learning.
Authors
Affiliations (5)
Affiliations (5)
- Graduate School of Science and Technology, Meijo University, 1-501 Shiogamaguchi, Tempaku-ku, Nagoya, 468-8502, Aichi, Japan.
- Graduate School of Science and Technology, Meijo University, 1-501 Shiogamaguchi, Tempaku-ku, Nagoya, 468-8502, Aichi, Japan. [email protected].
- Ibaraki Children's Hospital, 3-3-1 Futabadai, Mito, 311-4145, Ibaraki, Japan.
- Department of Radiology, Tokyo Metropolitan Children's Medical Center, 2-8-29,Musashidai, Tokyo, 183-8561, Fuchu, Japan.
- Faculty of Engineering, Gifu University, 1-1,Yanagido, Gifu, 501-1194, Japan.
Abstract
Children's bones are more elastic and have a thicker periosteum than those of adults, resulting in subtle, incomplete fractures. Diagnosis based on plain radiographs alone can be challenging. The shortage of pediatric radiologists compounds this difficulty, leading to a risk of missed fractures. Therefore, we considered that it might be possible to help prevent missed fractures by developing and validating an automated detection model for pediatric forearm fractures on plain radiographs using deep learning techniques (convolutional neural networks [CNNs] and Vision Transformers [ViTs]). To train and validate such a model, this study targeted the frontal and lateral views of plain radiographs from 517 patients aged 1-14 years with forearm fractures. We first performed preprocessing to focus on the forearm region in the images. Thirteen models (visual geometry group [VGG], ResNet, DenseNet, and ViT) were used to classify the presence or absence of fractures and were evaluated and compared using 5-fold cross-validation. Verification of these models showed that VGG16 exhibited the best performance. The overall result, which integrated the predictions from the frontal and lateral views, achieved a sensitivity of 0.872 ± 0.015, a specificity of 0.925 ± 0.016, a balanced accuracy of 0.898 ± 0.003, and an AUC of 0.962 ± 0.002. Furthermore, visualization of saliency maps (using gradient-weighted class activation mapping and an attention map) revealed that the model focused on the bone during prediction. Therefore, if the proposed method is used in hospitals, it could possibly support diagnosis and help reduce missed fractures.