Enhancing knee osteoarthritis diagnosis with AI: A deep learning perspective

Abstract

Knee Osteoarthritis is one of the most concerning diseases in the current world. A lot of people are suffering from Knee Osteoarthritis which is a common disease now-a-days. In the current world, the number of knee patients is increasing day by day. The risk of KOA increases for anyone who gets older or carries excess weights of body. Stress also causes KOA. Additionally genetic problems or joint injuries can cause KOA. Knee Osteoarthritis are two types. In the primary stage(KL-1 and KL-2) it is called Osteopenia and in the second stage(KL-3 and KL-4) is called Osteoporosis are significant global health concerns, leading to increased bone fragility and fracture risk. Usually patients feel pain in and around the knee, pain is frequently described as a stabbing, sharp, or dull ache. Effective clinical intervention depends on early and precise detection using X-ray imaging. However, manual diagnosis is difficult and subject to inter-observer variability due to minute differences in bone density and texture. In this paper, a new, lightweight architecture for a convolutional neural network is proposed, termed Res-SE LiteNet, designed to classify knee radiograph images into three classes: Normal, Osteopenia, and Osteoporosis. In order to minimize computational parameters, thereby making the model practical for a clinical environment, the proposed architecture utilizes Depthwise Separable Convolutions. To enhance feature sensitivity, we implemented a Channel-wise Attention Mechanism via Squeeze-and-Excitation (SE) blocks and Residual Skip Connections to prevent the degradation of low-level spatial details. A Hybrid Global Pooling strategy, combining Global Average and Max Pooling, was employed to capture both global bone density statistics and focal pathological markers. A dataset publicly available in Mendeley was used to train and validate the proposed model. The model was able to attain a peak validation accuracy of 90.53%. Compared to conventional models such as the VGG16 or ResNet50 models, the model is able to maintain high accuracy while using a much smaller number of parameters. By comparing computationally complex pre-trained models, our proposed lightweight model achieves comparable diagnostic accuracy with a significantly reduced parameter footprint. As a result the model can facilitate high-speed processing without compromising model accuracy.