Using Deep Networks for Knee Range of Motion Monitoring in Total Knee Arthroplasty Rehabilitation

Yixiang Zhang^1,2Qiu-yu Wang³Tao Yang^1,2Jia-he Wang⁴Haotian Yin^2,3Lei Wang²Jun Liu^1,2,3*

• ¹天津医科大学骨科临床学院, 天津市 / 市辖区 / 和平区, China
• ²Tianjin University Tianjin Hospital, Tianjin, China
• ³Tianjin University, Tianjin, China
• ⁴Tiangong University, Tianjin, China

Background: Knee range of motion (ROM) is a key indicator of rehabilitation after total knee arthroplasty (TKA). Current tools, such as visual and protractor measurements, are cumbersome, imprecise, and require professional training, limiting their use in community or home settings. With the rise of smart healthcare, there is a need for a simple, accurate, and low-cost ROM assessment method that reduces healthcare burdens, enables home self-monitoring, and improves rehabilitation outcomes. Methods: A total of 1,103 knee images were collected from 1,790 patients who had undergone TKA. The images were classified into four categories: standard flexion, substandard flexion, standard extension, and substandard extension and six categories: 0°, 25°, 50°, 75°, 100°, and 125°. The images were processed using KROMNet, which was trained with a deep learning architecture that included convolutional, dilated convolution, channel attention layers, and fully connected layers. The model was trained and evaluated using a dataset split into training and test sets, and its performance was assessed with precision, recall, F1-score, and accuracy metrics for both the four-class and six-class tasks. Results: KROMNet achieved an accuracy of 95.02% in the four-class task and 94.12% in the more challenging six-class task. In the four-class task, the precision, recall, and F1-score were 95.04%, 94.96%, and 94.98%, respectively. In the six-class task, KROMNet demonstrated an accuracy of 94.12%, with precision, recall, and F1-scores of 94.64%, 94.59%, and 94.60%, respectively. The model's performance was compared with other state-of-the-art methods, including Hazra's, Du's, Xia's, Victoria's, and Shiwei Liu's models, with KROMNet consistently outperforming these models in both four-class and six-class tasks. Conclusions: The KROMNet model proposed in this study offers an accurate, efficient, cost-effective, and remotely deployable solution for monitoring knee ROM after TKA. KROMNet not only demonstrates superior recognition performance under small sample conditions but also shows strong clinical utility and potential for wider adoption, making it especially suitable for grassroots, community, and home rehabilitation settings. KROMNet is expected to become a key tool in the intelligent rehabilitation system, helping healthcare reduce costs, increase efficiency, and improve patient experience and rehabilitation quality.