Recognition of Knee Osteoarthritis by 1D and 2D Convolutional Neural Networks Using Vibroarthrographic Signals

Osteoarthritis (OA) of the knee is a leading cause of joint pain and mobility loss. Early diagnosis is crucial for effective management yet remains challenging with traditional imaging techniques such as X-rays, computed tomography, and MRI. This study utilized 1D and 2D convolutional neural networks (CNN) to assess OA of the knee using vibroarthrographic (VAG) signals recorded by an inertial measurement unit sensor. VAG signals were recorded by deploying accelerometers at three key points on the knee joint, alongside knee movement data captured by an angle sensor. These signals were further processed, transforming them into time-domain features for 1D CNN and frequency-domain features for 2D CNN. The classification models—a depthwise separable 1D CNN and a pre-trained 2D MobileNet—aimed to distinguish between healthy individuals and those diagnosed with OA. Validation with 128 subjects yielded promising classification results. The 1D CNN demonstrated a strong benefit in the realizable edge computing system on raw VAG data, the memory size of only 234 Kbytes, while the 2D CNN, utilizing spectrogram images, achieved even higher classification accuracy of $0.764~\pm ~0.010$ . This work offers a low-cost, non-invasive, and effective method for knee OA diagnosis, potentially enhancing clinical assessments of joint disorders for outpatients.