Arthroscopy - 2026-09-06 - Journal Article
KneeFusionNet Enables Accurate and Efficient Comprehensive Detection of Knee Ligament Injuries on Magnetic Resonance Imaging: A Multicenter Validation Study.
Chen B, Tang X, Guo D, Shen X, Xu S, Li S, Li Y, Wang Q, Wu Y, Qin Y
Topics
Key Takeaway
KneeFusionNet achieved external validation AUROCs of 0.888 (ACL), 0.867 (PCL), 0.862 (MCL), and 0.874 (LCL) for knee ligament injury detection on MRI, while AI assistance improved junior surgeon accuracy from 0.818 to 0.900 and reduced interpretation time by 14.73 seconds.
Summary Depth
Choose how much analysis to show on this article page.
Summary
This 3-center retrospective study developed and externally validated KneeFusionNet, a DenseNet-based multimodal deep learning model, for simultaneous detection of ACL, PCL, MCL, and LCL injuries on knee MRI using arthroscopically confirmed cases as ground truth. Multimodal fusion outperformed all single-modality approaches and three comparison deep learning models on internal validation (all P<.05). External AUROCs ranged from 0.862 to 0.888 across the four ligaments, and AI assistance raised junior surgeon diagnostic accuracy by 8.2 percentage points while cutting interpretation time for all surgeons.
Key Limitation
The external test set comprised only 160 patients from a single center, limiting generalizability across diverse imaging protocols, scanner field strengths, and patient populations.
Original Abstract
PURPOSE
To develop and externally validate KneeFusionNet, a multimodal deep learning model for detecting anterior cruciate ligament (ACL), posterior cruciate ligament (PCL), medial collateral ligament (MCL), and lateral collateral ligament (LCL) injuries on knee magnetic resonance imaging (MRI), and to assess the impact of multimodal fusion and artificial intelligence (AI) assistance on diagnostic performance.
METHODS
This 3-center retrospective study was conducted between April 2020 and August 2025. The injury group included patients who underwent knee MRI within 1 month before arthroscopy and had surgically confirmed ACL, PCL, MCL, or LCL injuries; controls had unremarkable MRI and physical examination findings. Two centers formed the development set, and the remaining center served as the external test set. DenseNet-based KneeFusionNet was developed and compared with 3 deep learning models. Diagnostic performance was assessed using the area under the receiver operating characteristic curve, and a reader study evaluated AI-assisted diagnostic performance.
RESULTS
Overall, 919 patients were included: 759 in the development set and 160 in the external test set. Multimodal fusion outperformed single-modality approaches for all ligaments (all P < .05). On internal validation, KneeFusionNet achieved area under the receiver operating characteristic curves of 0.971 for ACL, 0.906 for PCL, 0.919 for MCL, and 0.924 for LCL. Corresponding external area under the receiver operating characteristic curves were 0.888, 0.867, 0.862, and 0.874. Sex-stratified analyses showed no consistent sex-related decrease in model performance. KneeFusionNet outperformed all comparison models on internal validation (all P < .05). AI assistance improved mean diagnostic accuracy for junior surgeons from 0.818 to 0.900 and reduced mean interpretation time by 14.73 seconds across all surgeons (all P < .05).
CONCLUSIONS
KneeFusionNet detected ACL, PCL, MCL, and LCL injuries on MRI with high diagnostic performance and outperformed comparison models. AI assistance improved diagnostic accuracy for junior surgeons and reduced interpretation time for all surgeons.
LEVEL OF EVIDENCE
Level III, retrospective case-control study.