ESTABLISHING A CONSECUTIVE CLINICAL KNEE OSTEOARTHRITIS IMAGING COHORT USING ARTIFICIAL INTELLIGENCE TOOLS TO ANALYZE KNEE RADIOGRAPHS: A PROOF OF PRINCIPLE

Traditional knee OA research cohorts are curated to test research hypotheses in controlled environments, but they lack near-clinical prevalence, current trends, and diversity in gene pools from around the world. Recently, artificial intelligence (AI) models have been shown to diagnose and grade the degree of knee OA in agreement with experienced consultants in musculoskeletal radiology. As a proof of concept, this study aims to create a scalable consecutive clinical knee OA imaging cohort that reflects current trends with limited use of human resources by using various AI tools. We included patients aged 35 to 79 with a weight-bearing knee radiograph from the University of Copenhagen Hospital, Bispebjerg-Frederiksberg Hospital, from January 1, 2016, to December 31, 2021. Four AI tools were applied to the retrospective cohort: 1) CE-marked and FDA-approved knee OA grading tool provided KL grades and minimal joint space width (MJSW) on bilateral frontal radiographs and the presence of patella osteophytes on lateral radiographs, 2) inhouse custom-build convolution neural network (CNN) radiographic view tool supported the OA grading tool confirming the inclusion of correct views, 3) inhouse custom-build CNN knee joint implant tool helped us exclude radiographs with implants, and 4) laterality marker detection tool detected missing or inconsistent markers on bilateral frontal images. We found 32,707 knee radiographs from 8,641 patients in the PACS and ended up with 25,854 bilateral frontal and lateral radiographs from 8,618 patients and their 9,689 examinations after the cleaning process with the four tools. Most knees (92.5 %) had complete AI-generated OA data (KL grade, MJSW, patella osteophyte results). We used no radiologist reading time, and only 16.0 % of the images were read by humans during quality control processes. It would have taken a radiologist about 800 hours to grade the radiographs if they had spent five minutes per examination. In this study, we showed that a consecutive clinical large-scale knee OA imaging cohort can be created from raw clinical images from the PACS system using a combination of four AI tools and no radiologist readings. This approach will enable us to test previous and new OA hypotheses on broader real-life clinical updated datasets that could help us diversify the findings and prevent data dredging. In addition, our approach can be used to create similar scalable cohorts from other geographic areas. In the future, we plan t