Thomas McKee Design Engineering Book a call →
Case study · Research

Predicting a fracture from the bone itself, using a scan the patient already had.

A research study bringing structural engineering to a clinical question: can a patient-specific finite element model, built from a routine CT scan, predict how likely a femur is to fracture? I carried out the structural analysis, owning the automated segmentation, meshing and convergence, material mapping, the boundary conditions including one of my own design, the load sweep, and turning strain into a clinical number.

Start your project → See the process
1317 N
average predicted minimum failure load, inside the osteoporotic range
−4.5%
failure load from my boundary condition, closer to the accurate benchmark
27
load cases: nine angles across three boundary conditions
0
invasive measures: no implant, biopsy or cadaver needed
The design process

How the study ran, end to end

A femur is a structure with complex geometry and properties that vary from point to point, under a known load. That makes it a finite element problem, and I approached it exactly as I would a performance part in mechanical design.

  1. Frame the clinical question as an engineering one

    What load does this particular bone fail at in a sideways fall? That's a question FEA can answer for one patient, where statistics can only answer it for a population.

    Problem definition · research

  2. Build the geometry from the scan

    An automated script segmented bone from non-bone slice by slice across several scans and planes, producing an STL surface for ANSYS.

    Scripting · image segmentation

  3. Mesh it and prove the mesh

    A convergence study across five element sizes found the point where the results stopped changing, so the mesh was fine enough to trust and coarse enough to run dozens of cases.

    Meshing · convergence study

  4. Give every element its own stiffness

    Material properties were mapped element by element from calibrated CT greyscale, so thousands of elements each carried a realistic stiffness.

    Material mapping · calibration

  5. Question the standard boundary conditions

    The published sideways-fall conditions restrain the knee in a way a real knee doesn't, so I designed and ran a third condition alongside them.

    Boundary condition design · load cases

  6. Turn strain into a clinical number

    A maximum principal strain criterion converted the results into a predicted failure load, benchmarked against cadaver experiments and the patient's bone density score.

    Post-processing · failure criteria · validation
Skills used

What this project called on

Engineering
FEAStructural analysisMesh convergenceMaterial mappingBoundary condition designFailure criteriaPost-processingValidation against experimentResearch
Tools
ANSYS MechanicalScripted CT segmentationCalibrated CT greyscaleSTL processing
Theory
Density-modulus relationshipsPrincipal strain failure criterionQuadratic tetrahedral elementsMulti-point constraintsContact modelling
Case study 5.1 · Clinical FEA

Predicting fracture likelihood from the bone's structure rather than from the patient's age

Problem

Fracture risk is normally judged from statistical factors such as age, weight and a bone mineral density score. Those are population-level correlations applied to one person, and they're a common cause of unnecessary surgery. The patient in front of the clinician either has a strong femur or doesn't, and a curve fitted to thousands of other people can't say which.

The clinical prize is a non-invasive, patient-specific measure. Everything in this study comes from a CT scan the patient has already had, with no implant, no biopsy and no cadaver.

What makes it hard: geometry from noisy greyscale data · material properties that vary element to element · a sideways-fall load case with no agreed boundary condition · a failure criterion that has to be a clinical number · a computing budget that has to allow dozens of load cases.
Fig. 78, The statistical approach being replaced: fracture probability against age
Fig. 79, The only input required: a clinical CT scan
Process

Segmentation, mesh convergence, a stiffness for every element, and a new boundary condition

  1. 01Automate the segmentation. Each CT frame separates bone from non-bone by greyscale. I wrote a script to segment slice by slice, ran it across multiple scans and planes and averaged the result automatically, which removed both the labour and the operator variation of doing it by hand. The output was an STL surface loaded into ANSYS Mechanical.
  2. 02Prove the mesh. To run many load cases the simulation had to be as efficient as possible, so I ran a convergence study across five element sizes using ten-node tetrahedra, tracking peak principal strain. Results were stable from four to three millimetres, moving around 2%, so a 2 mm mesh went forward with margin for the coarser regions. Quadratic elements were a must here: the geometry can't be swept or mapped, and linear tetrahedra badly under-predict strain in bending.
  3. 03Map the material element by element. Treating bone as one uniform material isn't accurate enough, so I set each element's properties from the calibrated greyscale. Density comes from a linear function of Hounsfield value, ρ = a + b·HU, calibrated against a reference phantom scanned alongside the patient, and stiffness from a published relationship, E = cρn. The phantom is what makes it quantitative, tying greyscale to real density for this scan on this machine.
  4. 04Question the standard boundary conditions. The published sideways-fall conditions, even the detailed ones, stop the knee rotating about its axis. A real knee isn't a torsional restraint, so locking it adds a reaction moment that doesn't exist in a fall, and any restraint that shouldn't be there makes a structure look stronger than it is.
  5. 05Design a better one. Alongside the two standard conditions I designed a third: a multi-point constraint at the knee with axial rotation free through ±30°. It costs the same to run as the detailed standard condition, with behaviour approaching the accurate but expensive contact-based benchmark.
  6. 06Sweep the load cases. Nine loading angles across all three conditions gave 27 cases in total, which is exactly why the convergence work mattered.
  7. 07Turn it into a clinical metric. A maximum principal strain failure criterion converted strain into a predicted minimum failure load, which I benchmarked against published cadaver experiments and the patient's own bone density score.
Fig. 80, Automated segmentation across planes
Fig. 82, Convergence study, peak principal strain
Fig. 83, Ten-node tetrahedral mesh at 2 mm
Fig. 84, Greyscale-to-density calibration against the phantom
Fig. 86, Boundary condition models used in the literature
Fig. 87, Nine-angle load sweep across all conditions
Results

Turning simulation into a clinical prediction

1317 Naverage predicted failure load
1 to 2.5 kNcadaver failure range
−4.5%from the new condition

The average predicted minimum failure load of 1317 N sits inside the osteoporotic range and at the low end of experimental cadaver failure loads, which run from 1,000 to 2,500 N. That can then be turned into a clinical prediction of the likelihood of fracture or refracture. Compressive strain proved the most likely failure mode, especially in falls.

My boundary condition predicted failure loads up to 4.5% lower than the standard ones, moving the result toward the far more accurate contact-based method without its computing cost. The results were validated against in vitro and ex vitro cadaver testing, and the research is continuing because of how well the initial studies went.

This involved some of the most technical set-up, meshing, boundary conditions and post-processing an engineer is likely to take on, and it's the same discipline I bring to a pressed oven cavity or a moulded tray: quantify the physics, validate it, then let the analysis carry the decision.

Fig. 88, Principal strain field, worst-case angle
Fig. 89, Healthy femur against osteoporotic femur, strain results
How I could help you

The same process, on your product

Simulation is only as good as its mesh, its materials and its boundary conditions. Setting those up properly, and checking them against reality, is what makes the results worth acting on. It's the core of Design Verification.

Let's talk

Need an analysis you can trust?

Book a free intro call and tell me what you need to know. I'll be honest about whether simulation is the right tool and how I'd set it up.

Book a free intro call → See more work