ACL Injury Assessment with AI: MRI to Decision Support
Deep learning for ACL tear detection from knee MRI — automated assessment, severity grading, surgical outcome prediction, and clinical deployment.
The Scale of the Problem
Anterior cruciate ligament (ACL) injuries affect an estimated 200,000 people annually in the United States alone, with the majority occurring in athletes between 15 and 45 years of age. Diagnosis relies on clinical examination (Lachman test, pivot shift, anterior drawer) confirmed by MRI, followed by a treatment decision that balances surgical reconstruction against conservative management based on the patient's age, activity level, and associated injuries.
MRI interpretation for ACL pathology is highly accurate in expert hands — sensitivity exceeding 95% for complete tears. But in practice, interpretation quality varies. Not every facility has a fellowship-trained musculoskeletal radiologist, and the time from MRI acquisition to final report can delay clinical decision-making by days. AI-assisted ACL assessment aims to close both the accuracy gap and the time gap.
What AI Can Detect in Knee MRI
Deep learning models for ACL assessment operate on two levels.
Binary detection answers the fundamental question: is the ACL intact or torn? Models trained on sagittal knee MRI sequences learn to recognise the normal low-signal band of the ACL running from the lateral femoral condyle to the anterior tibial plateau, and to flag its absence, discontinuity, or abnormal signal intensity.
The landmark research dataset for this task is Stanford's MRNet, containing 1,370 knee MRI exams with expert labels for three conditions: abnormality (any pathology), ACL tear, and meniscal tear. MRNet provided the first large-scale benchmark for knee MRI AI and demonstrated that deep learning models could achieve AUC (area under the ROC curve) values above 0.95 for ACL tear detection — approaching expert radiologist performance.
Severity grading goes beyond binary detection to characterise the injury: partial versus complete tear, location (proximal, midsubstance, distal avulsion), chronicity (acute oedema versus chronic scarring), and associated injuries (meniscal tears, bone bruise pattern, collateral ligament involvement). This multi-task assessment is more challenging and requires models that can process volumetric 3D MRI data rather than individual slices.
The Technical Architecture
ACL detection from MRI differs fundamentally from OA grading from radiographs. The input is volumetric: a knee MRI study consists of multiple sequences (sagittal T2, coronal PD, axial) each containing 20–40 slices. The model must integrate information across slices to assess a 3D structure (the ACL) from 2D sections.
2D approaches pass each sagittal slice independently through a CNN and aggregate predictions across slices — typically using max-pooling or a learned attention mechanism to identify the most informative slices. This is computationally efficient and was the approach used in the original MRNet paper.
3D approaches process the entire MRI volume as a 3D input, using 3D convolutional layers or transformer architectures that can capture spatial relationships between slices. These models better capture the continuous anatomy of the ACL but require more training data and computational resources.
Multi-sequence fusion combines information from sagittal, coronal, and axial sequences. Each sequence provides complementary information — sagittal views show the ACL fibre continuity, coronal views reveal the tibial and femoral attachment sites, and axial views can demonstrate the cross-sectional integrity. Models that cross-reference multiple sequences produce the most comprehensive assessment.
From Detection to Clinical Decision Support
Detecting an ACL tear is necessary but not sufficient for clinical decision-making. The surgeon needs to know not just whether the ACL is torn, but whether the patient will benefit from surgical reconstruction — a decision that depends on factors the MRI alone cannot fully answer.
Predicting surgical candidacy requires integrating imaging findings with clinical data: patient age, activity level, instability symptoms, associated meniscal or chondral injuries, and the patient's willingness to modify activity. Machine learning models trained on surgical registry data can estimate the probability of successful return to sport after reconstruction versus conservative management, helping inform the shared decision-making conversation.
Predicting graft choice and tunnel positioning for ACL reconstruction is an emerging application. Patient-specific femoral and tibial anatomy varies considerably, and AI models trained on post-operative outcome data can suggest optimal tunnel placement and graft sizing based on pre-operative imaging — bridging the gap between diagnostic imaging and surgical planning.
The Data Landscape
MRNet remains the primary public benchmark but has limitations. Studies were acquired at a single institution with consistent protocols on similar MR equipment. Models trained solely on MRNet frequently show performance degradation when applied to MRI data from other institutions — a domain shift problem that is common in medical imaging AI.
Beyond MRNet, available datasets include the fastMRI knee dataset (for reconstruction tasks rather than diagnosis), SKM-TEA (with cartilage lesion annotations), and various institutional collections published alongside individual research papers. There is no equivalent of the OAI dataset for knee MRI — a large, multi-institutional, longitudinal MRI dataset with comprehensive annotations would dramatically accelerate the field.
At Salnus, our dataset inventory includes MRNet (1,370 cases with ACL/meniscus/abnormality labels in NPY format, axial/coronal/sagittal sequences). This forms the starting point for our Motor 2 (MRI) development pipeline, which will complement our existing Motor 1 (X-ray) pipeline for knee OA assessment.
Clinical Workflow Integration
For ACL AI to reach clinical adoption, it must integrate into the existing radiology and orthopaedic workflow rather than creating a parallel process.
The ideal integration point is within the DICOM viewing environment. The surgeon or radiologist opens the knee MRI study in their viewer, and the AI model automatically processes the sequences in the background — producing an overlay or structured report that highlights the ACL status, any associated injuries, and confidence scores. This follows the same client-side inference architecture that we use for X-ray analysis: all processing occurs locally in the browser, and no patient data leaves the clinician's device.
GradCAM-style heatmaps adapted for MRI can highlight which slices and regions most influenced the model's prediction, giving the clinician an interpretable basis for accepting or overriding the AI assessment.
The Path Forward
ACL AI is at an earlier stage than knee OA AI — comparable to where OA grading was 3–4 years ago. Binary detection is solved at the research level but not yet externally validated for clinical deployment. Severity grading and outcome prediction are active research areas with promising but preliminary results.
For Salnus, ACL assessment represents our second AI motor — expanding from X-ray-based OA analysis to MRI-based ligament and meniscal assessment. This aligns with our vision of building a comprehensive orthopaedic AI platform that supports the surgeon across the full spectrum of knee pathology.
We are actively seeking sports medicine surgeons and MRI-focused radiologists interested in clinical collaboration — contributing anonymised MRI data, validating AI models against expert assessment, or co-developing clinical decision support tools. If this aligns with your research interests, contact our team.
Disclaimer: This article is for educational and research purposes only. AI tools referenced are designated for research use only (RUO) and are not cleared medical devices. Clinical decisions regarding ACL management should be made by qualified physicians based on comprehensive clinical and imaging assessment.
References:
- Bien N, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging: Development and retrospective validation of MRNet. PLoS Med. 2018;15(11):e1002699.
- Chang PD, et al. Deep learning for detection of complete anterior cruciate ligament tear. J Digit Imaging. 2019;32(6):980-986.
- Liu F, et al. Deep learning approach for evaluating knee MR images: achieving high diagnostic performance for cartilage lesion detection. Radiology. 2018;289(1):160-169.
Reviewed by the Salnus biomedical engineering team.