SWE-Net: Sliced-Wasserstein EfficientNet for Learning Discriminative Texture Features in Early Diagnosis of Knee Osteoarthritis

Osteoarthritis (OA) is a common degenerative joint disease that affects the cartilage and bone of the knee. Early diagnosis of OA can prevent further joint damage and improve patients' quality of life. In this paper, we propose a novel deep-learning approach for detecting the presence of OA in both the tibia and femur from plain radiographs. First, we used the Fast-SAM model to segment the knee joint areas to study the presence of OA within the femoral and tibial compartments. Then, we introduced a new Sliced-Wasserstein EfficientNet (SWE-Net) to promote discriminability and texture analysis in the feature space of the EfficientNet network. Specifically, we introduced a Sliced Wasserstein Texture (SWT) loss to break down high-dimensional distributions of the feature space into several one-dimensional marginal distributions. With this SWT loss, we capture higher-order statistics representing the texture feature distributions, helping to distinguish between similar classes, such as non-OA and early OA. Comprehensive experimental results over a large public database demonstrate the potential of the proposed network and facilitate a deeper understanding of localized pathological features associated with OA.