Deep learning methods to analyse knee osteoarthritis using magnetic resonance imaging: a systematic review and synthesis without meta-analysis

This systematic review focuses on understanding the evidence of validity of deep learning (DL) methods utilized to analyze knee osteoarthritis (OA) disease using Magnetic Resonance Imaging (MRI). A search was made in three scientific databases (Web of Science, PubMed, and Scopus) from January-1990 to March-2025 with the following keywords: Knee, Osteoarthritis, Deep Learning, Automatic, and MRI. Studies based on the key string search, inclusion, and exclusion criteria emphasizing the assessment and synthesis of DL methods using MRI for knee OA were only considered. Categorization of the outcome is based on five problem areas addressed: segmentation, classification, synthesis, registration, and outlier detection. A total of 56 studies were selected and critically assessed based on the problems associated with knee MRI for knee OA evaluation. The segmentation problem was addressed in 39 studies, classification in 14, registration in two, and synthesis in six studies. In addition, four studies simultaneously addressed segmentation and classification problems, while one study combined registration and segmentation tasks. Findings were synthesized using a structured synthesis without meta-analysis, informed by Cochrane guidance for reviews. Segmentation continues to be the predominant area of study, followed by classification. Most studies utilized a specific MR sequence and a specific dataset thereby leading to data-dependent DL methods. There is insufficient literature reporting unsupervised, interpretable, and explainable DL methods. In conclusion, standardization of validity and introduction of reliability measures is essential to ensure trustworthiness of these DL methods.