A Comparison of Transfer Learning Metaphyseal Sign Diagnostic Models for Kashin-Beck Disease Based on X-rays of Children’s Hands

Ge Li, Yujiao Zhang, Ping Chen, Yunlong Song, Yuchen Zhang, Chao LiCureus 2025 · DOI

Abstract

Background: Kashin-Beck disease (KBD), prevalent in certain regions of the world, primarily affects children and is characterized by joint deformities. Timely screening and accurate diagnosis, heavily reliant on metaphyseal signs in X-rays, are crucial but challenging, especially in regions where specialist availability is scarce. Artificial intelligence (AI)-assisted diagnostic technology offers a valuable solution to streamline KBD screening, emphasizing its importance in enhancing diagnostic precision and efficiency. Methods: This study developed and compared five deep learning models - KBV16, KBX, KBV19, KBIn, and KBM2 - to assist in diagnosing KBD by analyzing pediatric hand radiographs. We optimized these models with a dataset comprising 22,366 images, encompassing both metaphyseal positive and control groups. The models were trained and validated using Binary Cross-Entropy (BCE) and Accuracy (ACC) metrics. Results: The KBV16 model outperformed the others, achieving an accuracy of 0.9563 on the validation set and 0.9535 on the test set. The implementation of data augmentation techniques, along with the meticulous selection of learning rates and batch sizes, significantly enhanced the models' performance. Conclusion: This study presented a novel application of deep learning in KBD diagnosis, demonstrating the potential of AI models to enhance diagnostic precision. Notably, the KBV16 model emerged as a powerful tool for early detection of KBD. Future research should concentrate on refining these models for clinical use and integrating them into existing healthcare systems to improve medical services, particularly in medical resource-constrained regions.

Introduction

Kashin-Beck Disease (KBD) is an endemic osteoarthropathy that primarily affects children and adolescents in specific regions of Asia, most notably in China. The disease is characterized by the degeneration and necrosis of growth plate cartilage and articular cartilage, leading to severe joint deformities, muscle atrophy, and stunted growth. Early detection is paramount to managing the disease and preventing lifelong disability.

Diagnosis typically relies on the identification of "metaphyseal signs" in hand X-rays—subtle structural disruptions or depressions in the bone's calcification zone. However, interpreting these radiographs requires specialized radiological expertise, which is often unavailable in the remote, resource-constrained rural areas where KBD is most prevalent. This study explores the use of deep learning and transfer learning to automate the detection of these signs, providing a scalable screening tool for underserved populations.

Methodology

The researchers approached KBD diagnosis as a binary classification task, distinguishing between KBD-positive metaphyseal signs and healthy controls. To overcome the challenges of training deep models on limited medical data, the study employed transfer learning using five established convolutional neural network (CNN) architectures: VGG16 (KBV16), Xception (KBX), VGG19 (KBV19), Inception (KBIn), and MobileNetV2 (KBM2).

Dataset and Preprocessing

The dataset consisted of 22,366 hand X-ray images collected from surveillance facilities in Shaanxi province, China, focusing on children aged 7 to 12. The original positive class included 1,224 subjects. To address class imbalance, the positive samples were augmented 10-fold through random cropping, flipping, and adjustments to brightness and contrast, resulting in 12,240 positive images. The control group consisted of 10,126 images. The data was split into training, validation, and test sets using a 2:1:1 ratio.

Model Architecture and Optimization

The models were initialized with weights pre-trained on ImageNet. The final layers were replaced with trainable layers specific to the KBD classification task. Training utilized Binary Cross-Entropy (BCE) loss and the Adam optimizer. A critical component of the methodology was hyperparameter tuning, specifically focusing on the interaction between Learning Rate (LR) and Batch Size (BS).

Architecture of the KBV16 model based on VGG16.

Results

The comparative analysis revealed significant performance differences between the architectures. The KBV16 model, based on the VGG16 architecture, demonstrated the highest diagnostic accuracy.

Performance Metrics

KBV16 achieved a validation accuracy of 95.63% and a test accuracy of 95.35%. In contrast, other models showed lower performance: KBV19 (93.43%), KBX (90.54%), KBIn (86.69%), and KBM2 (82.84%). The superior performance of KBV16 suggests that its specific depth and convolutional structure are particularly well-suited for capturing the fine-grained textures and subtle irregularities of bone lesions in KBD.

Hyperparameter Optimization

The study found that the choice of hyperparameters was vital for model convergence. For the KBV16 model, the optimal configuration was a learning rate of 0.001 and a batch size of 16.

Validation accuracy across different learning rates and batch sizes for the KBV16 model.

Discussion and Conclusion

The success of the KBV16 model highlights the potential of AI to assist in the early diagnosis of KBD. While more modern architectures like Xception or MobileNetV2 offer computational efficiencies, they appeared less effective at identifying the specific metaphyseal markers required for KBD diagnosis. This suggests that for certain medical imaging tasks, traditional deep convolutional layers may be more robust than depthwise separable convolutions.

Limitations and Future Directions

Despite the high accuracy, the study has limitations. The data is geographically localized to Shaanxi province, and the labels are based on human interpretation, which may introduce bias. Additionally, the current models only perform binary classification and do not yet distinguish between different stages of disease severity.

Future work should focus on developing multi-class models for disease staging and integrating these algorithms into lightweight mobile applications. Such tools could empower local healthcare workers in remote regions to conduct rapid, on-site screenings, significantly improving the medical outlook for children at risk of Kashin-Beck Disease.

Summary updated

Abstract

Background Kashin-Beck disease (KBD), prevalent in certain regions of the world, primarily affects children and is characterized by joint deformities. Timely screening and accurate diagnosis, heavily reliant on metaphyseal signs in X-rays, are crucial but challenging, especially in regions where specialist availability is scarce. Artificial intelligence (AI)-assisted diagnostic technology offers a valuable solution to streamline KBD screening, emphasizing its importance in enhancing diagnostic precision and efficiency. Methods This study developed and compared five deep learning models - KBV16, KBX, KBV19, KBIn, and KBM2 - to assist in diagnosing KBD by analyzing pediatric hand radiographs. We optimized these models with a dataset comprising 22,366 images, encompassing both metaphyseal positive and control groups. The models were trained and validated using Binary Cross-Entropy (BCE) and Accuracy (ACC) metrics. Results The KBV16 model outperformed the others, achieving an accuracy of 0.9563 on the validation set and 0.9535 on the test set. The implementation of data augmentation techniques, along with the meticulous selection of learning rates and batch sizes, significantly enhanced the models' performance. Conclusion This study presented a novel application of deep learning in KBD diagnosis, demonstrating the potential of AI models to enhance diagnostic precision. Notably, the KBV16 model emerged as a powerful tool for early detection of KBD. Future research should concentrate on refining these models for clinical use and integrating them into existing healthcare systems to improve medical services, particularly in medical resource-constrained regions.
BibTeX
@inproceedings{li2025comparison,
  author = {Ge Li and Yujiao Zhang and Ping Chen and Yunlong Song and Yuchen Zhang and Chao Li},
  title = {{A Comparison of Transfer Learning Metaphyseal Sign Diagnostic Models for Kashin-Beck Disease Based on X-rays of Children’s Hands}},
  booktitle = {Cureus},
  year = {2025},
  doi = {10.7759/cureus.78235},
}
Figure