Deep semi-supervised learning for knee osteoarthritis severity assessment from plain radiographs

Abstract

Purpose: Deep Learning (DL) methods based on Supervised Learning (SL) have been recently adapted to assess knee osteoarthritis (OA) severity from plain radiographs. However, SL methods require large amounts of annotated data given by radiologists. While it is time-consuming and costly to obtain large annotated datasets, unannotated data are much easier to acquire at a low cost. In this study, we propose to use Semi-Supervised Learning (SSL) to leverage unannotated data (plain knee radiographs) in a combination with small amount of annotated data for automatic knee OA severity assessment.