Deep Learning for Real-Time Image Interpretation in Arthroscopy: A Systematic Review of Classification, Segmentation, and Detection Tasks

Background: Deep learning (DL) is increasingly applied to arthroscopic video for instrument recognition, anatomical segmentation, and intraoperative event detection. The performance, architectures, and real-time feasibility of these systems remain incompletely characterized. This systematic review aimed to evaluate the performance, architectures, and real-time feasibility of DL for classification, segmentation, and detection in arthroscopic video interpretation. Methods: Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020, PubMed/ MEDLINE, Ovid, and the Cochrane Library were searched (February 2025), supplemented by artificial intelligence-assisted queries (Elicit, Undermind). Eligible studies applied DL to arthroscopy for classification, segmentation, or detection with quantitative performance and/or inference speed. Two reviewers independently screened, extracted, and assessed quality using Quality Assessment of Diagnostic Accuracy Studies-2; the protocol was registered in International Prospective Register of Systematic Reviews. Results: Six studies met inclusion criteria (knee = 2, shoulder = 2, wrist = 1, non-site-specific bleeding = 1). Classification achieved high accuracy for instrument recognition (up to 99.1%, sensitivity = 0.988) and bleeding- frame detection (95.8%, sensitivity = 97.3%). Segmentation was strongest for large bony structures (Dice 0.79-0.89) and weakest for deformable tissues (meniscus Dice = 0.48). U-Net variants (U-Net++, DeepLab v3+) predominated with frequent transfer learning. Two studies achieved real-time inference (15-60 frames per second). Risk of bias was high across all studies. Conclusion: Based on 6 heterogeneous studies, DL appears technically feasible for arthroscopic image interpretation, particularly instrument classification and osseous segmentation, though findings should be interpreted as preliminary. No study evaluated impact on surgical decision-making or patient outcomes, and none was deployed prospectively in live surgery. Translation will require multicentre datasets, standardized reporting, and prospective evaluation against clinically meaningful endpoints. Cite this article as: Lasslett C, Reddy R, Giritharan G, et al. Deep learning for real-time image interpretation in arthroscopy: a systematic review of classification, segmentation, and detection tasks. Eurasian J Med. 2026, 58(4), 1415, doi: 10.5152/ eurasianjmed.2026.261415.