Residual Network with Triple-Attention Mechanisms for Knee Osteoarthritis Severity Classification

¹ Systems science, School of Mathematics and Statistics, Chongqing Jiaotong University, China
² Fuling Center for Disease Control and Prevention, Chongqing, China

^* Corresponding author: Rqy2077@outlook.com

Abstract

As the quality of life continues to improve, people are more concerned about all types of diseas- es. Knee osteoarthritis (KOA) is a type of arthritis that is characterized by limited movement, joint stiffness and pain. This degenerative disease leads to gradual wear and tear of the knee joint and in severe cases, dis- ability. Conventional radiographic diagnosis remains challenging due to the subtle morphological changes in early-stage KOA that often resemble age-related physiological variations. Meanwhile, applying convolu- tional neural networks to the prediction of KOA has become an effective method. To address this diagnostic bottleneck, we introduce Triplet Attention (TA) in Residual Network (ResNet) for KOA recognition. The architecture innovatively integrates cross-dimensional attention modules within residual blocks, enabling simultaneous modeling of channel-wise dependencies, spatial correlations, and hierarchical feature relation- ships. Our experimental framework was rigorously evaluated on two medical imaging datasets, comprising 8,263 standardized knee radiographs with Kellgren-Lawrence grading annotations. The proposed architec- ture demonstrated statistically improvements over other existing convolutional networks (VGG-19, Dense- Net-121, GoogleNet, etc.).