Meniscus Tear Detection with AI: Why Detection Outpaces Localisation in Knee MRI
AI detects meniscal tears with 87% sensitivity — but accurately localising them remains a challenge. How deep learning is advancing meniscus assessment, where the gaps are, and what clinicians should know about the detection-localisation asymmetry.
The Meniscus: A Unique Challenge for AI
The menisci are among the most commonly injured structures in the knee — and among the most difficult for AI to assess accurately. Unlike ACL tears, which present as a disrupted linear structure on MRI, meniscal tears are morphologically diverse: horizontal cleavage tears, radial tears, complex tears, bucket-handle tears, and root tears each have distinct appearances and distinct surgical implications.
This morphological diversity makes meniscal assessment one of the harder problems in musculoskeletal radiology — and one where AI performance has historically lagged behind human experts.
Prevalence of degenerative changes. In patients over 40, meniscal signal abnormalities are extremely common — many represent asymptomatic degenerative changes rather than clinically significant tears. An AI model that flags every signal abnormality will have high sensitivity but poor specificity, generating excessive false positives that erode clinical trust.
What Does the Evidence Show?
Systematic Reviews and Meta-Analyses
A 2024 systematic review and meta-analysis in the European Journal of Radiology examined AI performance for meniscal tear detection across published studies. The pooled results revealed an important asymmetry: AI was better at detecting the presence of tears (sensitivity 87%, specificity 89%) than at accurately locating them within specific anatomical sub-regions (sensitivity 88%, specificity lower and more variable).
This finding has direct clinical implications. Knowing that a meniscal tear exists is useful for screening, but surgical planning requires knowing where it is. AI that detects but cannot localise offers limited value to the treating surgeon.
A 2025 systematic review published in Cureus examined AI-based meniscal tear detection across multiple studies, confirming that binary detection performance is approaching clinical utility — but location-specific classification remains less reliable. A separate 2025 meta-analysis in PLoS One reinforced these findings across a broader dataset.
Technical Approaches
MRNet (Bien et al., 2018) remains a foundational benchmark. This Stanford dataset of 1,370 knee MRI examinations with labels for ACL tears, meniscal tears, and general abnormalities established the deep learning approach to knee MRI interpretation. Published models achieve AUC of 0.84–0.87 for meniscal tear detection on this dataset.
nnU-Net-based approaches that combine segmentation with classification have shown improved performance. Using nnU-Net for segmentation followed by 3D CNN classification, AUC values of 0.89–0.93 have been reported.
YOLOv8 + EfficientNetV2 (Gungor et al., 2025) achieved mAP@50 of 0.98 for meniscus localisation and AUC of 0.97–0.98 for tear classification — the highest published results. However, this study used 642 knees with expert annotation of specifically selected slices (one sagittal, one coronal per patient), which may overestimate performance on full volumetric data.
Weakly supervised approaches (FAD-MIL, 2026) address the annotation bottleneck by learning from image-level labels rather than pixel-level segmentation, achieving AUC 0.833 on the FracAtlas-equivalent meniscus dataset — lower than fully supervised approaches but requiring dramatically less annotation effort.
The Location Problem
The most clinically relevant finding across the literature is that AI performs substantially better at binary detection (tear present/absent) than at anatomical localisation (which compartment, which horn, which tear pattern).
This matters because treatment decisions depend on localisation. A small radial tear in the red-red zone (peripheral vascularised region) may be amenable to repair, while the same tear in the white-white zone (avascular central region) is more likely to require partial meniscectomy. A bucket-handle tear needs urgent surgical attention; a stable horizontal cleavage tear may be managed conservatively.
Until AI can reliably distinguish these patterns, its role in meniscal assessment remains primarily as a screening tool — flagging cases that warrant careful radiologist review — rather than a diagnostic tool that can inform surgical planning directly.
Multi-Structure Assessment: The Future Direction
Beyond Single-Structure Models
The knee does not tear one structure at a time. ACL ruptures are frequently associated with lateral meniscal tears. PCL injuries often accompany complex multi-ligament patterns. Meniscal tears commonly coexist with cartilage damage and early osteoarthritic changes.
Multi-task models that simultaneously assess ligament, meniscus, cartilage, and bone from a single MRI acquisition represent the natural evolution of knee AI. These models mirror the radiologist's approach — evaluating the entire joint rather than isolated structures — and are more computationally efficient than running separate models for each pathology.
The MRNet dataset includes labels for ACL tears, meniscal tears, and general abnormalities, making it a natural platform for multi-task learning research. Published multi-task models have shown that joint learning across structures improves performance for each individual task compared to single-task models.
Integration with Bone Segmentation
Meniscal assessment does not exist in isolation. Combining soft tissue evaluation (meniscus, cartilage, ligaments) with bone segmentation and geometric analysis creates a comprehensive joint assessment platform.
For example, a medial meniscal tear combined with varus mechanical axis alignment (measured from automated geometric analysis) carries a different prognosis — and may warrant different treatment — than the same tear in a neutrally aligned knee. Integrating these assessments into a unified platform supports more nuanced clinical decision-making.
What Clinicians Should Know
AI is a screening tool for meniscal tears, not a diagnostic replacement. Pooled sensitivity of 87% and specificity of 89% is useful but insufficient for stand-alone diagnosis.
Detection outpaces localisation. AI can reliably tell you that a meniscal tear exists. It is less reliable at telling you exactly where it is and what type it is.
MRI sequence matters. Most published models are trained on specific sequences (proton density-weighted, T2-weighted). Performance may degrade on sequences the model has not seen.
Degenerative signal is the main source of false positives. In older patients, AI models frequently flag grade 2 intrasubstance signal as tears. Understanding this limitation prevents unnecessary surgical referrals.
Multi-structure models are coming. Rather than separate tools for ACL, meniscus, and cartilage, expect integrated platforms that assess the entire knee joint from a single MRI — matching the radiologist's holistic approach.
Explainability is particularly important for meniscal assessment. Because tears are morphologically diverse and can be subtle, the ability to see which region of the MRI the model is attending to is critical for trust calibration.
Where Salnus Fits
Salnus's MRI analysis pipeline (Motor 2) is currently focused on ACL assessment, with meniscal tear detection as a planned extension. The Stanford MRNet dataset — which we use for ACL model development — includes meniscal tear labels, providing a natural path for multi-task model development.
Our bone segmentation pipeline, achieving Dice scores of 0.964 across four knee bones, complements soft tissue assessment by providing the geometric context needed for comprehensive joint evaluation. The combination of bone geometry, ligament integrity (ACL, PCL), and meniscal status represents the long-term vision: a unified AI-powered knee assessment platform.
For clinicians and researchers interested in meniscal AI assessment or multi-structure knee MRI analysis, contact our team →
Disclaimer: This article is for educational and research purposes only. Salnus tools are designated for Research Use Only (RUO) and are not cleared medical devices. Clinical decisions regarding meniscal pathology should be made by qualified orthopaedic surgeons based on comprehensive clinical and imaging assessment.
References:
- Zhao Y, et al. AI applied to MRI reliably detects the presence, but not the location, of meniscus tears: systematic review and meta-analysis. Eur Radiol. 2025;35(5):2952-2953.
- Giammanco PA, et al. Diagnostic Accuracy of AI for Detection of Meniscus Pathology on MRI: A Systematic Review. Cureus. 2025;17(9):e91832.
- Gungor N, et al. High accuracy in meniscus tear detection using YOLOv8 and EfficientNetV2. KSSTA. 2025.
- Bien N, et al. Deep-learning-assisted diagnosis for knee MRI: MRNet. PLoS Med. 2018;15(11):e1002699.
- Tack A, et al. Multi-Task Deep Learning for Detection of Meniscal Tears. Frontiers Bioeng Biotechnol. 2021;9:747217.
- Mohammadi S, et al. Diagnosis of knee meniscal injuries using AI: systematic review and meta-analysis. PLoS One. 2025;20(6):e0326339.
Reviewed by the Salnus biomedical engineering team.