Comprehensive Summary
This study introduces Artificial Intelligence for Knee Imaging Registration and Analysis (AKIRA), a deep learning (DL)-based tool for automated classification and annotation of knee radiographs in patients with anterior cruciate ligament (ACL) injuries. The system aims to improve radiographic standardization, implant detection, and osteoarthritis (OA) grading for the development of a comprehensive ACL injury registry. The dataset consisted of 20,836 knee radiographs from 1,628 ACL-injured patients, with a median age of 26 years (IQR: 19–42), 57% male, and an average follow-up of 70.7 months (IQR: 6.8–172 months). AKIRA integrates three deep learning models: EfficientNet for laterality and projection classification, YOLO for implant detection, and a residual network for OA grading based on Kellgren–Lawrence (KL) classification. The laterality and projection model achieved F1 scores between 0.941 and 1.0, indicating high accuracy in identifying radiographic projections (anterior- posterior [AP], lateral, sunrise, posterior-anterior [PA], hip–knee–ankle [HKA], and Camp- Coventry intercondylar notch views). The implant detection model demonstrated an area under the precision-recall curve (AUPRC) of 0.695–0.992, successfully identifying 16,112 ACL-related implants, including femoral metal screws (5,210), femoral metal buttons (3,563), tibial metal screws (3,780), and tibial metal buttons (989). The OA classification model exhibited concordances between 0.39 and 0.40 for KL grading, improving to 0.81–0.82 for binary OA classification (KL grade ≥2). AKIRA successfully processed and labeled all 20,836 images within 88 minutes, enabling large- scale, AI-assisted radiographic registry creation for ACL injury management. The study highlights the potential of AI-driven imaging analysis in standardizing diagnostics, improving research capabilities, and facilitating personalized treatment planning.
Outcomes and Implications
The implementation of AKIRA presents a transformative approach to radiographic standardization and AI-assisted diagnostics in musculoskeletal imaging. The system significantly improves the efficiency, accuracy, and scalability of ACL injury registries, allowing for automated classification of knee radio graphs based on laterality, projection, implant presence, and OA severity. The ability to analyze 20,836 images in under 90 minutes represents a substantial improvement over traditional manual labeling processes, which are prone to human error and inconsistencies. The implant detection model’s high precision (AUPRC: 0.695–0.992) allows for reliable identification of ACL fixation devices, supporting surgical planning and postoperative monitoring. Similarly, the OA grading model’s binary classification accuracy (0.81–0.82) provides a clinically relevant method for tracking long-term joint health and degenerative changes in ACL-injured patients. By integrating AI into clinical workflows, AKIRA has the potential to improve orthopedic research, enhance patient outcomes, and streamline diagnostic processes. Technology could facilitate personalized treatment strategies by identifying early signs of post-ACL reconstruction OA, ultimately guiding preventive care and rehabilitation protocols. Future directions should focus on expanding AKIRA’s capabilities to include real-time analysis, integration with electronic health records, and cross-institutional validation studies. Further optimization of OA grading accuracy and implant detection in diverse patient populations could enhance its clinical applicability. Multimodal AI models incorporating MRI and CT scans may further improve diagnostic precision, leading to more effective, AI-driven orthopedic care. The success of AKIRA underscores the growing role of AI in musculoskeletal imaging, paving the way for automated radiographic analysis, improved registry development, and data-driven decision-making in orthopedic surgery.