AI in Osteotomy Planning: Where the Field Stands and Where It's Going
A comprehensive analysis of AI applications in corrective osteotomy — from automated angle measurement to end-to-end surgical decision support systems.
Corrective osteotomy around the knee — whether high tibial (HTO), distal femoral (DFO), or tibial tubercle — remains one of the most planning-intensive procedures in orthopaedic surgery. The difference between a good outcome and a revision often comes down to a few degrees: the correction angle, the hinge point, the tibial slope you preserve or intentionally alter. For decades, that precision has depended entirely on the surgeon's eye, a goniometer, and experience.
Artificial intelligence is beginning to change this equation — but not in the way most people assume. The current generation of AI tools in osteotomy is narrowly focused on measurement automation. The next generation, which several groups including ours at Salnus are actively working toward, aims at something far more ambitious: end-to-end surgical decision support, from the initial radiograph to the final correction plan.
This post maps the landscape as it stands today, identifies the gaps that matter most, and outlines where the field is heading.
Where We Are: AI as a Measurement Tool
The most mature AI applications in osteotomy planning are automated radiographic measurement systems. These tools use deep learning — typically convolutional neural networks trained on annotated radiographs — to detect anatomical landmarks and compute angles that surgeons currently measure by hand.
Mechanical axis and alignment angles are the most common targets. Several groups have demonstrated AI systems that automatically extract the hip-knee-ankle angle (HKA), medial proximal tibial angle (MPTA), lateral distal femoral angle (LDFA), and joint line convergence angle from standing long-leg radiographs. These systems consistently achieve accuracy within 1–2° of expert manual measurement, often with superior reproducibility.
Posterior tibial slope (PTS) has emerged as a particularly active area of research. The slope, measured from lateral knee radiographs, is critical in osteotomy planning because HTO inherently increases the posterior slope — an effect that has implications for ACL strain, knee kinematics, and long-term outcomes. AI tools that automatically measure PTS reduce interobserver variability and — critically — enable standardised data collection across multi-centre studies.
But they solve only the first step of the planning workflow.
What's Missing: The Decision Support Gap
Here is the core problem: automated measurement is not the same as automated planning.
A surgeon planning an HTO does not simply need to know the MPTA is 82°. They need to integrate that number with the LDFA, the overall mechanical axis deviation, the joint line obliquity, the patient's weight and activity level, the state of the cartilage in each compartment, and — increasingly — the tibial slope they intend to produce. The correction target is not a fixed formula; it depends on whether the goal is neutral alignment, slight undercorrection, or Fujisawa-point targeting based on the cartilage status.
Current AI tools cannot do this. They measure angles. They do not recommend correction targets, simulate osteotomy wedge geometry, or evaluate the biomechanical consequences of a planned correction. Two patients with identical MPTA values may need different corrections if one has an intact lateral compartment and the other shows early lateral degeneration. This requires integrating imaging data with clinical context — something no current automated system does.
The ESSKA consensus on osteotomy around the degenerative varus knee has formalised many of these decision pathways into structured clinical frameworks. But translating those expert-driven rules into software that a surgeon can interact with at the point of planning remains an unsolved problem.
Where It's Going: End-to-End Planning Systems
The next generation of AI in osteotomy planning will need to move beyond single-measurement tools toward integrated planning systems. We see this happening in three stages:
Stage 1 — Automated comprehensive radiographic analysis. Rather than measuring one angle at a time, a single pipeline that accepts a standing long-leg AP and a lateral radiograph and returns a complete biomechanical profile: mechanical axis, MPTA, LDFA, JLCA, PTS, and derived parameters like mechanical axis deviation in millimetres. This is largely achievable with current technology and removes the most time-consuming part of manual planning.
Stage 2 — Correction target recommendation with simulation. Given the complete biomechanical profile plus clinical inputs (cartilage status, patient factors, surgical goals), the system recommends a correction angle and hinge point, simulates the resulting alignment, and predicts the post-operative tibial slope. The surgeon reviews, adjusts, and confirms. This requires not just measurement AI but biomechanical modelling and clinical decision logic.
Stage 3 — 3D-guided surgical execution. The confirmed correction plan maps onto patient-specific 3D anatomy from CT, generating cutting guides (PSI), osteotomy plane visualisation, and intraoperative navigation parameters. This closes the loop from diagnostic image to surgical execution, with AI providing quality checks at each stage.
Some elements of Stage 3 already exist in isolation. AI-based 3D preoperative planning systems for total knee arthroplasty — such as the G-NET system published in Scientific Reports — demonstrate that neural networks can identify tibial and femoral landmarks from CT, simulate osteotomy planes, and predict prosthesis sizing with over 90% accuracy. Adapting these approaches to corrective osteotomy, where the geometry is more variable and the correction targets are patient-specific, is a natural but technically demanding next step.
The Regulatory and Practical Reality
It would be incomplete to discuss AI in surgical planning without acknowledging the regulatory landscape. Any AI system that recommends a surgical correction — as opposed to simply measuring an angle — is functioning as clinical decision support software and falls under medical device regulations (EU MDR Class IIa/b, FDA Class II). This is appropriate. A system that influences how much bone a surgeon removes must meet a higher evidence bar than one that measures an angle on a radiograph.
The practical consequence is that the most clinically impactful AI applications — Stages 2 and 3 above — face the longest path to clinical deployment. This creates a tension: the tools surgeons need most are the ones that take longest to validate. The teams that navigate this effectively will be those that build their AI on a foundation of published clinical evidence, rather than attempting to shortcut the validation process.
Where Salnus Fits
Our work at Salnus sits at the intersection of these trends. Our cloud-based platform delivers AI-powered surgical planning tools directly to the surgeon's browser — no PACS integration, no software installation. Our platform already supports DICOM viewing, AI-powered bone segmentation with 0.964 mean Dice score, and structured measurement workflows. The next steps involve exactly the kind of integrated planning pipeline described above: from automated measurement through decision support to 3D-guided execution.
We believe the teams that will succeed in this space are those that combine deep clinical knowledge with production-grade engineering — and that publish their work openly rather than hiding behind marketing claims. That is the standard we hold ourselves to.
Key Takeaways
- AI in osteotomy planning today is primarily limited to automated angle measurement, which is valuable but insufficient for surgical planning.
- The critical gap is decision support: translating measurements into actionable correction plans that account for deformity origin, slope effects, and patient-specific factors.
- The field is moving toward integrated end-to-end systems that span from radiographic analysis to 3D-guided surgical execution.
- Regulatory requirements for decision-support software are substantial and appropriate — clinical validation is not optional.
- At Salnus, we are building toward this vision with a foundation of peer-reviewed research, cloud-native engineering, and close collaboration with orthopaedic surgeons.
Salnus is a biomedical AI company focused on orthopaedic surgery. Learn more at salnus.com.
Reviewed by the Salnus biomedical engineering team.