AI and 3D reconstruction of the knee joint from MRI is transforming preoperative planning as a radiation-free and acurate alternative to CT scans.
The information in this article is drawn from a peer-reviewed study by Prof. Cavaignac and his peers: “Three-dimensional reconstruction of the knee joint based on automated 1.5T magnetic resonance image segmentation: A feasibility study“.
Generating precise 3D models of the knee used to require a CT scan. Today, artificial intelligence combined with standard MRI technology is changing that: faster, safer and without radiation.
This technology is primarily aimed at sports medicine and knee ligament surgery, with ACL reconstruction as the main target. It’s particularly relevant for young, active patients where avoiding radiation from CT scans is an important consideration.
CT Scan 3D Reconstruction of the Knee: Why It’s No Longer Enough
CT Scan 3D reconstruction of the knee has long been the reference standard for surgical planning. It produces detailed images of bony structures, converted into 3D models surgeons use to prepare complex procedures.
But CT has real limitations:
- Ionising radiation: a concern for young athletes requiring repeated imaging
- Poor soft tissue visualisation: ligaments, menisci and cartilage remain largely invisible
- Limited accessibility: advanced CT equipment is not available everywhere
MRI addresses all three issues. It produces no radiation, captures soft tissues with excellent clarity and is already part of the standard workup for most knee injuries.
The remaining challenge? Turning MRI images into accurate 3D models quickly and reliably, and that’s exactly where AI steps in.
How AI Segmentation Makes MRI-Based 3D Knee Modelling Work
Creating a 3D model from an MRI scan requires segmentation which means identifying and outlining specific anatomical structures slice by slice. Traditionally done by hand, this process took nearly four hours per knee.
Three Approaches Tested
The study compared three methods:
| Method | Description | Average time |
| Manual (MRI MS) | Expert outlines each structure by hand | ~3.9 hours |
| Fully automated (MRI A) | AI performs segmentation alone | ~27 seconds |
| Semi-automated (MRI SA) | AI + expert review and correction | ~25 minutes |
The AI model used (UNet-R) is a transformer-based deep learning architecture trained to recognise 12 knee structures: bones, cartilage, menisci and ligaments. It processes each scan in approximately 22 seconds.
How Accurate Are the Results?
Accuracy was measured by comparing MRI-derived models to laser surface scanning (ground-truth precision: 30 microns). Key findings:
- Semi-automated: 1.05 mm (femur) / 1.03 mm (tibia); below the 1.5 mm clinical threshold
- Manual: 0.99 mm / 0.93 mm; most precise, but far more time-consuming
- Fully automated: 1.19 mm / 1.54 mm; fastest, but occasional errors requiring correction
The semi-automated approach offers the best balance between precision and efficiency.
Clinical Benefits of MRI-Based 3D Knee Reconstruction
MRI-based 3D knee reconstruction unlocks several concrete advantages for patients and surgeons alike.
No More Radiation
Removing CT from the preoperative workflow eliminates radiation exposure entirely, especially relevant for young patients or those needing multiple scans over time.
A Complete Picture of the Knee
MRI captures bone and soft tissue in a single scan. For ACL reconstruction, where associated meniscal or cartilage injuries are common, this is a major clinical advantage.
Accessible to More Hospitals
Most prior research relied on 3 Tesla MRI systems, which offer higher resolution but are less widely available. This study validates the workflow on 1.5 Tesla scanners, which are standard in most hospitals worldwide.
From Scan to Custom Surgical Instruments
MRI-derived 3D models can be used to manufacture personalised surgical instruments (PSI), tools designed around a patient’s exact anatomy, improving precision during surgery.
Limitations and What Comes Next
This study has important caveats worth noting:
- Small sample size (11 cadaveric knees): limits generalisability
- Cadaveric tissue vs. living tissue: the AI was trained on scans from living subjects; differences in signal may explain occasional segmentation errors
- No in vivo validation yet: real-world performance in living patients still needs to be confirmed
The next steps include larger multicentre studies, direct comparison with CT in living patients and integration with augmented reality guidance during arthroscopy (overlaying 3D models onto the surgical field in real time using SLAM algorithms).
Conclusion & Findings
MRI-based 3D reconstruction of the knee is no longer experimental. Sub-millimetre accuracy, a 25-minute workflow and zero radiation make it a credible, patient-friendly alternative to CT-based modelling, ready for broader clinical adoption.
To go further, we invite you to explore the other studies published by the research team involved in this work.
For patients seeking expert care in knee ligament surgery, Professor Étienne Cavaignac combines cutting-edge research with frontline clinical practice, making him an excellent reference for complex knee conditions.