Leveraging Deep Learning for the Automated Detection of Hemophilic Arthropathy in Knee Radiographs

Authors

En-Wei Lin, Department of Medical Education, Changhua Christian Hospital, Changhua, TaiwanFollow
Edzer Wu, DeepQ Technology Corporation, New Taipei City, TaiwanFollow
Ming-Ching Shen, Hemophilia Treatment and Thrombosis Center, Department of Internal Medicine, Changhua Christian Hospital, Changhua, Taiwan
Shao-Li Han, School of Chinese Medicine, College of Chinese Medicine, China Medical University, Taichung, Taiwan

Abstract

Background/Purpose:

Hemophilic arthropathy (HA) is a progressive joint disease caused by recurrent hemarthroses in patients with hemophilia. Early radiographic identification of HA is critical for timely intervention and long-term joint preservation. This study aimed to develop and evaluate deep learning models for automated detection of HA in knee radiographs.

Methods:

A retrospective dataset of 849 knee radiographs (AP and lateral views) was used to train three convolutional neural networks: ResNet-34, DenseNet-121, and EfficientNet-B0. Model performance was assessed using standard classification metrics, including accuracy, precision, recall, F1-score, and AUC. Interpretability was evaluated using Grad-CAM heatmaps.

Results:

DenseNet-121 achieved balanced performance on AP views (F1-score: 72.8%), while ResNet-34 performed best on lateral views (F1-score: 71.0%). Grad-CAM visualizations confirmed anatomically meaningful model attention. Misclassifications were most frequent in subtle early-stage HA or in normal variants with confounding features.

Conclusion:

This proof-of-concept study demonstrates the feasibility of deep learning–based detection of HA in knee radiographs. The proposed models may support image-based diagnosis in hemophilia care, especially in settings with limited radiology expertise.

Recommended Citation

Lin, En-Wei; Wu, Edzer; Shen, Ming-Ching; and Han, Shao-Li (2025) "Leveraging Deep Learning for the Automated Detection of Hemophilic Arthropathy in Knee Radiographs," Rehabilitation Practice and Science: Vol. 2026: Iss. 1, Article 2.
DOI: https://doi.org/10.6315/3005-3846.2272
Available at: https://rps.researchcommons.org/journal/vol2026/iss1/2