Effect of training set sample size on the agreement, accuracy, and sensitivity to change of automated U-net-based cartilage thickness analysis

Abstract

Purpose: Fully automated segmentation of articular cartilage from MRI using deep learning (DL), has shown promising results, with a recent focus on convolutional neural network (CNN) U-Net architectures. We have previously reported that models that have been trained on a specific radiographic disease stage (KLG) show superior automatic segmentation results compared with more general models that have been trained on knees with various radiographic stages. However, it is currently unknown, how the sample size of such (specific) training model affects the performance of the U-net-based segmentation and cartilage thickness analysis.