Quantifying Uncertainty in Knee Osteoarthritis Diagnosis

Knee OsteoArthritis (OA) is one of the most common causes of physical disability in the world, causing a large personal and socio-economic burden. Visual assessment of OA still suffers from subjectivity. Deep learning (DL), and in particular convolutional neural networks (CNN), has recently led to remarkable improvements in knee OA detection. However, traditional deep learning-based knee OA classification algorithms lack the ability to quantify decision uncertainty. This is a key point in the medical field where, due to the high cost of labelling, we are faced with a lack of sufficient data to train a learning model. We propose here an alternative approach based on the the concept of Evidential Deep Learning (EDL). Unlike Bayesian neural networks which indirectly infer prediction uncertainty through uncertainties in the network weights, EDL approaches explicitly model this uncertainty using the theory of subjective logic. Experimental results on the Osteoarthritis (OAI) database demonstrate the potential of the proposed approach.