DeepKneeXR: YOLOv8 multi-label X-rays detection of knee abnormalities from sports injury with clinical explainability

BACKGROUND: Soft-tissue knee abnormalities are common, yet first-line radiography provides limited soft-tissue contrast, whereas MRI or arthroscopy is more resource-intensive. We developed DeepKneeXR as a single-center, retrospective proof-of-concept AI workflow for generating probability scores for key knee abnormalities from anterior-posterior knee X-rays. METHODS: This retrospective study included 3,200 adult patients selected from 5,000 initially screened cases at one medical center after predefined exclusions. Reference labels were assigned using a composite clinical-imaging standard based on clinical history, physical examination, MRI findings, and arthroscopy when clinically indicated. A unified YOLOv8 model was trained to perform knee localization and multi-label probability prediction in a single forward pass. RESULTS: The model generated a knee bounding box and probability scores for meniscus tears (MENI), medial collateral ligament injuries (MCL), and joint effusion (EFFU). DeepKneeXR achieved excellent knee localization (mAP@0.5=0.995). Multi-label screening performance was moderate and should be interpreted as preliminary, particularly for EFFU, whose validation AUC was limited. CONCLUSIONS: This proof-of-concept study shows that a unified YOLOv8 model can generate knee-localization outputs and abnormality probability scores from AP radiographs. However, prospective multi-center validation, standardized reference labeling, and clinician-facing workflow evaluation are required before clinical use can be considered.