Deep Learning for Knee Osteoarthritis Detection and Severity Grading: A Comprehensive Study Using Radiographic Imaging and Grad-CAM Interpretability

Abstract

Knee Osteoarthritis (OA), like other forms of degenerative arthrosis, is characterised by progressive joint degeneration, the formation of bone spurs in the sub-chondral bone, and narrowing of the joint spaces and is one of the most disabling conditions worldwide. For the OA diagnosis, radiographic imaging is the most used and best standard method, and the Kellgren–Lawrence (KL) grading system is the most commonly used to assess severity. Unfortunately, manual grading takes a lot of time and is subject to inter-observer variability. For this gap, we created a deep learning supervised model for automated OA detection and severity classification, using the publicly available Knee Osteoarthritis Dataset with Severity Grading. For this, we trained several and varied architectures of convolutional neural networks (CNN), including VGG16 and MobileNet. For the model created, we used contrast normalisation, histogram equalisation, and image resizing to 224x224 pixels, followed with extensive data augmentation. The models created in this research are trained using transfer learning with weights from ImageNet, and the Adam optimiser with a categorical cross entropy loss. Moreover, for joint space narrowing and bone spurs, we used Grad-CAM for attention and alignment of the clinically relevant features to explain the decision of the model. Based on experimental outcomes, MobileNet exhibits the greatest balance between precision and efficiency, attaining an overall classification precision of 87.3%. With respect to the VGG16 and baseline models, the weighted F1 score and Cohen’s kappa were 0.85 and 0.81, respectively, signifying that MobileNet is the most accurate and reliable. Grad-CAM visualisations depict the integrating areas which correspond to the radiological features of OA, which increases the clinical trustworthiness. This deep learning study highlights the precision, interpretability and scalability available for the diagnosis and grading of the clinical and radiological features of knee OA, emphasising the technology’s potential to aid large-scale screening and clinical decision-making.

Ähnliche Arbeiten