Automated CT Bone Segmentation for Surgical Planning: How It Works and What to Expect

TL;DR

Automated CT bone segmentation uses deep-learning models to trace bone boundaries (femur, tibia, patella, pelvis) across a CT volume in minutes, a task that takes a technician 2 to 4 hours by hand. Modern models (nnU-Net, TotalSegmentator) reach accuracy approaching expert manual segmentation, but the output must remain editable: AI occasionally misses anatomy or includes artefacts, and the surgeon or technician verifies before the plan reaches the operating room. The practical value is not just speed, it makes 3D planning economically viable for routine cases, not only complex revisions.

Why Segmentation Is the Bottleneck

Every 3D orthopaedic plan starts the same way: turning a stack of grey CT slices into separate, labelled 3D bone models. This step, segmentation, is the foundation for everything downstream: alignment measurement, implant templating, osteotomy simulation, and patient-specific guides.

Done manually, it is slow and tedious. An experienced technician traces the contour of each bone across dozens or hundreds of slices, 2 to 4 hours for a single knee. That cost is why, for years, 3D planning was reserved for complex revision cases and PSI, never routine surgery.

How Deep-Learning Segmentation Works

A segmentation model learns to recognise bone boundaries from thousands of annotated CT volumes. Given a new scan, it predicts, for every voxel, which structure it belongs to: femur, tibia, patella, background.

The current reference architectures are well established:

• nnU-Net — a self-configuring framework that adapts its preprocessing and network to the dataset; the de facto benchmark for medical image segmentation.
• TotalSegmentator — a model trained to segment a broad set of anatomical structures from CT, validated across multiple regions.

These reduce a multi-hour manual task to minutes of compute, and the accuracy on bone is high, with overlap metrics (Dice scores) approaching expert manual segmentation for well-acquired CT.

What to Expect: Accuracy, Time, and the Editing Step

Accuracy. On clean, well-acquired CT, automated bone segmentation is reliable enough for clinical planning workflows. Accuracy drops with metal artefact (existing implants), severe deformity, or low-dose/noisy acquisitions, exactly the cases where a surgeon's review matters most.

Time. Minutes of inference versus hours of manual tracing. The economic consequence is the real story: routine cases become viable for 3D planning.

The editing step is not optional. Automated output will occasionally over- or under-segment, bridge a joint space, or include an artefact. A clinical-grade tool lets you correct the segmentation, combining AI speed with human verification. A tool that outputs a fixed, uneditable mesh should be treated as a research demo, not a planning instrument. (Comparison of 3D segmentation tools for hip planning discusses these accuracy/editability trade-offs.)

The Practical Workflow: DICOM to 3D Model

1. Load the DICOM series. A DICOM-native tool reads pixel spacing and slice thickness automatically, so the resulting model is physically calibrated.
2. Run segmentation. The model labels the bones; this is the minutes-not-hours step.
3. Review and edit. The surgeon or technician inspects the masks slice by slice and corrects errors.
4. Generate the 3D model. Labelled masks become surface meshes for measurement, templating, or PSI design.
5. Plan. Alignment angles, implant overlay, osteotomy simulation, all built on the verified model.

In the Salnus Surgeon Portal this runs client-side in the browser: DICOM parsing, segmentation, and rendering happen on the surgeon's device, so no patient imaging is uploaded to a server.

Where It Helps, and Its Limits

Automated segmentation reliably accelerates the foundation of 3D planning and reduces inter-observer variability in the downstream measurements. What it does not do is replace clinical judgment: the surgeon interprets the model, weighs patient-specific factors, and owns the surgical strategy. And a model trained on knee CT does not transfer to hip, shoulder, or spine, each anatomy needs its own training and validation.

FAQ

How accurate is automated CT bone segmentation? On clean CT, overlap with expert manual segmentation is high (Dice scores in the high 0.9s for large bones are common in the literature). Accuracy falls with metal artefact, severe deformity, or noisy scans, which is why editable output and surgeon review matter.

How long does it take? Minutes of compute, versus 2 to 4 hours for manual tracing of a single knee.

Can I edit the automated segmentation? You should be able to. Clinical-grade tools make the masks editable so you can correct missed anatomy or artefacts before planning.

Does my CT have to be uploaded to the cloud? Not with client-side tools. Browser-based platforms segment on your device with no upload, which simplifies KVKK/GDPR/HIPAA obligations.

The Takeaway

Automated CT segmentation turned 3D planning from a specialist, multi-hour process into a minutes-long, routine-viable step. The two things to check in any tool: is the accuracy validated for your anatomy, and is the output editable so you stay in control before the plan reaches the OR.

Explore the Salnus Surgeon Portal →

---

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. Mention of third-party tools is for educational context only. Clinical decisions should be made by qualified physicians.

References:

• Comparison of Three 3D Segmentation Software Tools for Hip Surgical Planning. PMC, 2022. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9323631/
• Isensee F, et al. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods, 2021.
• Wasserthal J, et al. TotalSegmentator: robust segmentation of anatomical structures in CT. Radiol Artif Intell, 2023.
• Salnus Research Group. 3D-Printed Patient-Specific Guides for Knee Reconstruction. OJSM, 2026.