AI in Orthopaedic Surgery: Where We Stand in 2026
Comprehensive guide to AI in orthopaedic surgery — radiographic OA grading, MRI ligament analysis, 3D surgical planning, robotics, regulatory landscape, and privacy architecture.
The Current Landscape
Artificial intelligence in orthopaedics has moved beyond proof-of-concept. As of 2026, deep learning models are actively deployed or in advanced clinical validation across four major domains: diagnostic imaging, surgical planning, implant design, and outcome prediction. The question is no longer "can AI help orthopaedic surgeons?" but rather "where does it add the most value today, and what should surgeons demand before trusting it?"
This article maps the current state across seven domains, names the key players, and provides a framework for evaluating tools — including an honest assessment of where Salnus fits.
Radiographic AI: OA Grading and Fracture Detection
Radiograph interpretation is where orthopaedic AI has made the most measurable progress. The fundamental value proposition is not accuracy but consistency. Two radiologists examining the same knee radiograph agree on the exact KL grade only 50–65% of the time. AI does not eliminate this variability, but it provides a reproducible baseline that anchors human assessment.
Performance benchmarks have converged around 80–95% accuracy for binary OA detection and 70–85% for five-class KL grading. The remaining accuracy gap concentrates in borderline cases — the KL grade 1 vs. 2 distinction where even experts disagree — rather than clear pathology.
Several commercial and research systems are active in this space:
Radiobotics (Copenhagen) has published external multi-site validation of their RBknee system, demonstrating that AI-assisted reading increases experienced orthopaedic surgeons' inter-reader agreement and accuracy. Their CE-marked tool is among the few with external validation data published in peer-reviewed journals.
KOALA (Knee Osteoarthritis Automatic Labelling Algorithm) uses ensemble methods to combine multiple CNN architectures, achieving robust performance across imaging protocols. The project has been validated on the OAI and MOST datasets.
MediAI-OA approaches the problem through multi-task learning — training a single model to simultaneously predict KL grade, joint space width, and osteophyte severity, mirroring the radiologist's approach of evaluating the entire joint rather than isolated structures.
Fracture detection has seen commercial deployment faster than any other orthopaedic AI application. Systems trained on datasets like FracAtlas (4,083 annotated radiographs in COCO and YOLO format) can identify fractures in the wrist, hip, and spine with sensitivity exceeding 90% in several published studies. The clinical utility is strongest in emergency departments, where AI serves as a second reader to catch fractures missed on initial assessment.
MRI Analysis: Ligament and Meniscus Assessment
MRI-based AI represents the next frontier. While radiographic AI works with single 2D images, MRI AI must process volumetric data — dozens of slices across multiple sequences — to assess three-dimensional structures like the ACL, PCL, and menisci.
ACL tear detection has been the most active research area. Stanford's MRNet study (published in PLoS Medicine, 2018) demonstrated that deep learning could detect ACL tears from knee MRI with AUC exceeding 0.96. Subsequent work has expanded to multi-ligament assessment and meniscal pathology.
More recent models combine 2D slice-level analysis with 3D volumetric processing, achieving robust performance across different MRI sequences (PD, T2, STIR). The clinical challenge remains generalisability — models trained on one scanner vendor or field strength may underperform on another.
The PCL — often called the "forgotten ligament" — presents particular challenges for AI. Its biomechanical complexity and the relative scarcity of isolated PCL injuries in training datasets make it a harder target than ACL detection. Multi-ligament models that assess all four major knee ligaments simultaneously represent the current research direction.
Surgical Planning and 3D Reconstruction
AI is accelerating the transition from generic to patient-specific surgical planning. Three applications are gaining traction.
Automated segmentation uses deep learning to extract bone surfaces from CT or MRI volumes, replacing hours of manual segmentation with minutes of automated processing. Models like nnU-Net and SAM-Med3D can segment femur, tibia, and patella from knee CT with accuracy approaching expert-level manual segmentation. This capability feeds directly into 3D-printed patient-specific surgical guides and preoperative planning platforms.
Geometric analysis automates the measurement of anatomical angles — mechanical axis (HKA), medial proximal tibial angle (MPTA), lateral distal femoral angle (LDFA), and femoral-tibial angle (FTA). These measurements are foundational for deformity assessment and correction planning (HTO, DFO). AI-based measurement reduces operator variability and enables rapid pre-surgical planning that would take significantly longer with manual methods.
Implant templating uses machine learning to predict optimal implant size and positioning from preoperative imaging, presenting data-driven outcome estimates alongside the surgeon's clinical experience and the patient's preferences.
The challenge is validation. Prediction models trained on one institution's data frequently underperform on another's — reflecting differences in patient populations, surgical techniques, and rehabilitation protocols. Multi-institutional validation is essential but logistically complex.
Outcome Prediction and Decision Support
Beyond imaging analysis, AI is being applied to clinical outcome prediction. Models trained on surgical registry data can estimate the probability of specific outcomes — return to sport after ACL reconstruction, revision risk after total knee arthroplasty, or functional improvement after HTO.
The main challenge is data quality: outcome prediction requires large, longitudinal datasets linking preoperative imaging and clinical data to postoperative results over months or years. Most existing datasets lack this temporal depth.
Predictive models are being developed to estimate functional outcomes after total knee arthroplasty, helping set realistic expectations and identify patients who may benefit from alternative treatment pathways.
Robotics and Intraoperative AI
Robotic-assisted orthopaedic surgery has grown significantly, with systems like MAKO (Stryker), ROSA (Zimmer Biomet), and VELYS (DePuy Synthes) deployed across thousands of centres worldwide. While these platforms incorporate computational planning, their AI sophistication varies.
Current robotic systems primarily execute pre-planned bone cuts with submillimetre precision — the "intelligence" is in the planning phase rather than real-time adaptation. The next generation of intraoperative AI aims for closed-loop feedback: systems that adjust surgical parameters in real time based on tissue properties, instrument position, and intraoperative imaging.
The integration of AI with surgical robotics is not just about precision — it is about augmenting the surgeon's decision-making during the procedure. Real-time feedback on soft tissue balance, ligament tension, and component alignment could transform outcomes in arthroplasty and deformity correction.
Regulatory Landscape
The regulatory environment for orthopaedic AI is evolving rapidly across jurisdictions.
EU MDR classifies clinical decision support software under Rule 11. AI tools that influence diagnostic or treatment decisions — the category most orthopaedic AI falls into — are generally classified as Class IIa or IIb, requiring Notified Body assessment and CE marking.
FDA has adopted a Total Product Lifecycle (TPLC) approach for AI-enabled devices, with draft guidance issued in January 2025. The Predetermined Change Control Plan (PCCP) framework allows manufacturers to pre-specify anticipated algorithm updates — critical for AI models that improve over time with additional training data.
Turkey (TİTCK) has harmonised with EU MDR, meaning CE-marked devices can access the Turkish market through the ÜTS registration system.
For surgeons evaluating AI tools, the key question is regulatory status. A CE-marked, clinically validated tool carries different weight than a research prototype — regardless of how impressive the demo appears. The distinction between "Research Use Only" and "cleared medical device" is not bureaucratic fine print; it reflects whether the tool has undergone external validation, safety assessment, and quality management scrutiny.
Privacy Architecture: A Differentiating Factor
How an AI tool processes patient data is as important as what it analyses. Two fundamental architectures exist:
Server-side processing sends medical images to external servers for AI analysis. This approach allows more powerful models (larger architectures, GPU inference) but creates data governance challenges — patient imaging data leaves the institution's network, requiring robust data processing agreements, encryption, and regulatory compliance.
Client-side inference runs AI models directly in the surgeon's browser using frameworks like ONNX Runtime Web. The patient's imaging data never leaves the hospital network — a significant advantage for KVKK/HIPAA compliance. At Salnus, our binary OA screening model (27MB, DenseNet-121) runs entirely client-side with approximately 110ms inference time per image.
The trade-off is model size: client-side inference requires smaller, optimised models that can run in a browser environment. Server-side systems can deploy larger models but must solve the data governance problem.
What Surgeons Should Look For
In 2026, five factors distinguish reliable tools from impressive demos:
Regulatory clearance. CE marking or FDA clearance indicates external validation. Research Use Only tools may be valuable for evaluation but should not inform clinical decisions.
External validation. Performance reported on the developer's own test set is necessary but insufficient. Multi-institutional, multi-vendor external validation provides stronger evidence of real-world reliability.
Explainability. Tools that show why they reached a conclusion — through GradCAM heatmaps, confidence scores, or highlighted anatomical features — support clinical trust and enable the surgeon to identify cases where AI assessment may be unreliable.
Integration. AI that requires exporting images, uploading to a separate platform, and manually entering results into the clinical record will not be used in practice. Integration into existing DICOM viewing workflows is essential for adoption.
Privacy architecture. Understanding where patient data flows — whether it stays on-premises or is transmitted to external servers — is a clinical governance responsibility, not just an IT concern.
Salnus's Position
Salnus is building an AI-powered clinical decision support platform for orthopaedic surgeons. Our current capabilities include: knee OA screening (DenseNet-121, 84.1% binary accuracy on the OAI test set), multi-planar reconstruction DICOM viewing, and automated clinical reporting.
Our platform uses client-side inference — all AI processing runs in the surgeon's browser, with no patient data transmitted to external servers. This architecture provides KVKK and HIPAA compliance by design.
We are currently at Research Use Only status, with development underway toward CE marking and TİTCK registration. Our approach prioritises transparent performance reporting, GradCAM explainability, and clinical validation through academic partnerships.
For surgeons interested in evaluating our platform or collaborating on clinical validation research, 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 should be made by qualified physicians based on comprehensive patient assessment.
References:
- Brejneboel MW, et al. External validation of an AI tool for radiographic knee osteoarthritis severity classification. Eur J Radiol. 2022;149:110194.
- Pihl K, et al. AI-based computer-aided system for knee osteoarthritis assessment increases experienced orthopaedic surgeons' agreement rate and accuracy. Knee Surg Sports Traumatol Arthrosc. 2023;31:1028-1039.
- Bien N, et al. Deep-learning-assisted diagnosis for knee magnetic resonance imaging. PLoS Med. 2018;15(11):e1002699.
- Mercurio M, et al. Deep Learning Models to Detect ACL Injury on MRI: A Comprehensive Review. Diagnostics. 2025;15(6):776.
- FDA. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. Draft Guidance, January 2025.
Reviewed by the Salnus biomedical engineering team.