Fully automated cartilage segmentation of late stage osteoarthritis (KLG4) knees using deep learning with U-net architectures - model specificity is important

Abstract

Purpose: The agreement and accuracy of automated cartilage segmentation of knees with radiographic osteoarthritis (ROA) using deep learning (DL) and convolutional neural network (CNN)-based U-Net architectures has recently been reported. It was observed that when using a model trained with a set of Kellgren Lawrence grade (KLG) 2-4 knees for this purpose, automated cartilage segmentation slightly overestimated cartilage thickness vs. the manual reference segmentation. This overestimation was particularly strong in a subcohort of KLG4 knees with denuded subchondral bone areas.