The following is a summary of the “KNee OsteoArthritis Prediction (KNOAP2020) challenge: An image analysis challenge to predict incident symptomatic radiographic knee osteoarthritis from MRI and X-ray images,” published in the January 2023 issue of Osteoarthritis and Cartilage by Hirvasniemi, et al.
Objectively comparing approaches for predicting the incidence of symptomatic radiographic knee osteoarthritis within 78 months on a test set with blinded ground truth was the goal of the KNee OsteoArthritis Prediction (KNOAP2020) challenge, which was designed to test these hypotheses. Participants in the challenge could use whatever data they deemed appropriate for training their models. All participants were given access to a test set of 423 knees from the Prevention of Knee Osteoarthritis in Overweight Females (PROOF) trial, which included magnetic resonance imaging (MRI) and X-ray image data, as well as clinical risk indicators at baseline.
Participants were not given access to the raw data, i.e., whose knees acquired incident symptomatic radiographic knee osteoarthritis (according to the combined ACR criteria) within 78 months. Researchers used both receiver operating characteristic curve (ROCAUC) and balanced accuracy (BACC). An impressive 23 entries were submitted by 7 different groups. Data collected by the Osteoarthritis Initiative was used to educate most of the algorithms. The best model for ROC AUC (0.64 (95% CI: 0.57-0.70)) uses deep learning to extract features from X-ray images and clinical data.
3 separate models were ensembled using automatically generated features from X-rays, MRIs, and clinical factors to produce the model with the highest BACC (0.59 (95% CI: 0.52-0.65). Predicting the incidence of radiographic knee osteoarthritis was given a standard by the KNOAP2020 challenge. More research is needed to overcome the difficult problem of accurately predicting incident symptomatic radiographic knee osteoarthritis.
Source: sciencedirect.com/science/article/pii/S1063458422008640