Which Uncertainty Quantification Method in Deep Learning Models for Knee Osteoarthritis Diagnosis?

This article focuses on uncertainty quantification (UQ) in deep learning methods for knee osteoarthritis (OA) diagnosis. Knee OA is one of the leading causes of physical disability worldwide, with a high personal and socio-economic burden. However, the detection of knee OA relies on visual assessment, which is highly subjective. Recently, deep learning has led to remarkable improvements in the detection of knee OA. However, the deep learning methods proposed in the current literature seem to be more concerned with model performance than with UQ.In this paper, we explore several UQ methods that aim to complement existing deep learning approaches. In particular, we consider conformal prediction, Monte Carlo dropout and evidential deep learning. The advantages and disadvantages of each method are highlighted in the context of the particular application. The aim is to provide a simple means of expressing uncertainty while allowing greater model transparency for clinical users.