Abstract
OBJECTIVE: Artificial intelligence (AI-) based automated cartilage analysis demonstrated similar sensitivity to change and only slighty inferior differentiation between radiographic progressors and non-progressors compared with manual segmentation. However, this finding was based on DESS MRI from the Osteoarthritis Initiative (OAI), whereas the vast majority of multicenter clinical trials rely on T1-weighted gradient echo (e.g. FLASH). Here we directly compare fully automated analysis of coronal FLASH vs. sagittal DESS, and vs. manually segmented DESS, in a sample with both FLASH and DESS MRI acquisitions. DESIGN: Convolutional neural network (CNN) algorithms were trained on 86 radiographically osteoarthritic knees with manual expert segmentation of the medial and lateral femorotibial cartilages (coronal FLASH and sagittal DESS). Post-processing involved automated registration of CNN-based subchondral bone segmentation to reference areas. The models were applied to baseline and two-year follow-up MRIs of radiographic progressor and non-progressor knees in the Foundation of the NIH Biomarker sample of the OAI. RESULTS: Of the 322 FNIH knees with both FLASH and DESS; 157 were radiographic progressors. Sensitivity to medial femorotibial cartilage thickness change (standardized response mean) in the progressor subcohort was -0.81 for FLASH (automated analysis), -0.74 for automatically, and -0.72 for manually segmented DESS. Differentiation from non-progressors (Cohen's D) was -0.82. -0.70, and -0.87, respectively. CONCLUSIONS: Fully automated, AI-based cartilage segmentation with advanced post-processing reveals that coronal FLASH is at least as discriminative between radiographic progressor vs. non-progressor knees as sagittal DESS MRI. Yet, performance of fully automated segmentation is slightly inferior to manual analysis with expert quality control. TRIAL ID: Clinicaltrials.gov identification: NCT00080171.