Lesson 1: Personalizing individual segments based on geometric data from medical images

This tutorial presumes that you have read the AMMR documentation and know how to personalize your model using information about height, weight, and individual segment lengths.

This lesson introduces an advanced approach to scaling based on a sequence of affine and non-affine transformations. Each of these transforms is constructed based either on subject-specific geometry or on a set of landmarks selected on the bone surface. As opposed to the simple scaling laws explained in the AMMR documentation, this lesson is rather methodological than conceptual and provides a good overview of how to pipeline and combine different 3D transforms to obtain subject-specific morphing and registration between frames of reference.

Linear point-based scaling

Scaling schemes described in the AMMR documentation are based on anthropometric measurements and affine transform scaling. Such schemes are good assumptions when more accurate measurements are not feasible or not available. Therefore, these schemes are used quite often. However, a natural next step would be to improve the precision of a model by utilizing subject-specific geometry available from the medical images. Medical images contain more subject-specific information about the bone shapes and local deformities that cannot be handled by the anthropometric regression equations.

The simplest inclusion of the subject-specific bone shape from medical image data is to find the affine (linear) transformation that fits a number of corresponding points that are selected on the source and the target geometries. These points could be fitted e.g. in a least-squares manner. This approach is similar to utilizing external body measurements as it relies on a linear transform. However, it is less dependent on the bone orientation and prior knowledge of dimensions to be measured. For example, you can locate any two points on source and target surface consistently without thinking of how a segment length was measured.

Let us make a simple example of using landmark-based affine scaling. First, please download two femur surfaces, SourceFemur.stl and TargetFemur.stl and save them in your working directory. These femur geometries will be used for the rest of this tutorial. The source surface is an unscaled femur used in the standard AnyBody models in the AMMR. The target surface is a femur reconstructed from a CT image and saved as a surface mesh in STL format (courtesy of Prof. Sebastian Dendorfer, OTH Regensburg, Germany).

Next, please download the AnyScript file lesson3a.main.any. This file contains a model with two segments which contain the definition of a surface each, one for the source (bone color) and one for the target (yellow) bone. When we load this model, the Model View should show the following picture:


To define a new scaling function let us insert a new AnyFunTransform3DLin2 object after the target segment:

AnyFunTransform3DLin2 <ObjectName> =
  //PreTransforms = {};
  Points0 = ;
  Points1 = ;

The AnyFunTransform3DLin2 object allows us to build a transform that fits a set of source and target landmarks in a least-squares manner as mentioned before. The object constructs a linear transforms in a full affine (linear transformation with translation, rotation, size-scaling and skewing, i.e. 12 degrees of freedom), uniform (orthogonal rotation with uniform scaling and translation, i.e., 9 d.o.f.), or rigid-body manner (orthogonal rotation of unscaled object with translation, i.e., 6 d.o.f.). Please note that the AnyFunTransform3DLin2 object utilizes the vtk-function/filter vtkLandmarkTransform, and, therefore inherits its modes:




A description of this function can be found here.

For this example we want to register our source surface into the target one by using a full affine transform. Therefore, we select several corresponding points on the surfaces and put them into the two point-sets called Points0 and Points1, which are the source and target points, respectively. As the next step, we change the mode of the AnyFunTransform3DLin2 object to VFK_LANDMARK_AFFINE to use the affine transform:

AnyFunTransform3DLin2 MyTransform =
  //PreTransforms = {};
  Points0 =
      {{-0.00906139,    0.36453,  0.0175591}, // fovea capitis
      {0.0358368,   -0.0100391, -0.0162062}, // lateral anterior condyle
      {0.0295267,   -0.0112881,  0.0194889}, // medial anterior condyle
      {0.0282045,     0.157599, -0.0172379}, // anterior mid shaft
      {-0.0245689, -0.00701566, -0.0238393}, // lateral posterior condyle
      {-0.0320739, -0.00877602,  0.0244234}};// medial posterior condyle

  Points1 =
      {{0.289913,0.420538,0.0138931},     // fovea capitis
      {0.322038,0.433232,-0.378636},    // lateral anterior condyle
      {0.289309,0.426839,-0.372994},    // medial anterior condyle
      {0.328859,0.425856,-0.175012},    // anterior mid shaft
      {0.306293,0.487243,-0.370319},    // lateral posterior condyle
      {0.261891,0.47585,-0.372696}};    // medial posterior condyle

The selected points on the surface represent specific anatomical landmarks and points described in the comments of the AnyScript code. Final modification before we can use the constructed linear transform is to give this transformation a name and apply it to the source surface:

AnySeg SourceFemur =
  Mass = 0; Jii = {0, 0, 0};
  AnyDrawSurf Surface =
    FileName = "SourceFemur.stl";
    RGB = {222,202,176}/256;
    AnyFunTransform3D &ref = ..MyTransform;

Reloading the model and looking at the bones shown in the Model View, we can see that these bones are now merged. To highlight the differences, we select one of them. This will produce the following picture.


The source bone is now transformed, i.e., translated, scaled and skewed to match the target bone. To make that clear, let us add a new AnyFunTransform3DLin2 called MyTransform2 to the model which we place after MyTransform. The intention is to construct a rigid-body registration transform between target and source surface. Please note, the roles of the source points Points0 and target points Points1 are swapped, and the transformation mode is set to VTK_LANDMARK_RIGIDBODY.

Additionally to that, a combination transform, containing forward affine and back registration transforms, is added:

AnyFunTransform3DLin2 MyTransform2 = {
  Points0 = .MyTransform.Points1;
  Points1 = .MyTransform.Points0;
AnyFunTransform3DIdentity MyTransform3 =
  PreTransforms = {&.MyTransform,&.MyTransform2};

Finally, let us look at the effect of the constructed transform. We comment the transform used in the visualization of the source surface and create another surface that will show the combined transformation that we just constructed:

AnySeg SourceFemur =
  Mass = 0; Jii = {0, 0, 0};
  AnyDrawSurf Surface =
    FileName = "SourceFemur.stl";
    RGB = {222,202,176}/256;
    //AnyFunTransform3D &ref = ..MyTransform;
  AnyDrawSurf SurfaceMorphed =
    FileName = "SourceFemur.stl";
    AnyFunTransform3D &ref = ..MyTransform3;

Looking at the Model View, we can see that the femur is now scaled, it became shorter and now aligns with the original source femur position. From the previous picture, we also know that geometry is matching the target quite well too (and if you want to convince yourself you can superimpose the target geometry using the MyTransform2 registration transformation).

With this example, we have shown how to morph the source into the target with a full affine scaling and subsequently applying a reverse registration to move the morphed geometry back.

Notice that it is possible to reverse the combination, i.e., to apply the registration step first and then the scaling/morphing step. For instance, make a transformation similar to MyTransform, but insert MyTransform2 as pre-transformation. In this tutorial lesson, we shall however stay with the concept we presented so far.

If the morphing accuracy is sufficient for your task you can proceed with your modeling and stop at this step. However, for the purpose of this tutorial the desired accuracy have not been reached - some local features still do not match the target’s ones, e.g. the lesser and the greater trochanter. The following steps explain how to capture more details and improved morphing for even better match.

Incorporating landmark-based nonlinearities into the scaling function

The next level of detail can be achieved by introducing local nonlinear deformations by means of the AnyFunTransform3DRBF class. This class represents a nonlinear interpolation/extrapolation transformation, which is based on the Radial Basis Functions (RBF) method and uses landmarks selected on source and target surfaces. Detailed behaviour of this transform is described in an appendix tutorial. However, the focus of this tutorial is to demonstrate available pipelines of transforms. For simplicity, we use a preselected set of femoral landmarks and RBF settings.

We start with the model from the previous steps to introduce the landmark-based nonlinear scaling. Several tranformations will build up into a pipeline, where pre-transforms will be used to inherit obtained accuracy throughout different steps. A complete model can you find here: lesson3b.Main.any. The following tutorial shows how to add an RBF transform with the recommended settings into the previously created model.

First of all let us configure the visualization of the transformation. Now that we know how to compare source and scaled geometries as well as reverse registration, so we can switch off the registration step.

AnyFunTransform3DIdentity MyTransform3 =
  PreTransforms = {&.MyTransform/*,&.MyTransform2*/};

This will return our morphed geometry back to the target bone location and we can observe the improvements as we go. Let us now define an RBF transformation and another AnyDrawSurf object that will show the difference between the affine scaling and the new transformation pipeline employing nonlinear RBF transformations. For a better contrast of the different surfaces, we will also add some colors to the drawing of the surfaces:

  AnyDrawSurf SurfaceMorphedRBF =
    FileName = "SourceFemur.stl";
    AnyFunTransform3D &ref = ..MyRBFTransform;


AnyFunTransform3DRBF MyRBFTransform =
  PreTransforms = {&.MyTransform};
  PolynomDegree = 1;
  RBFDef.Type = RBF_Triharmonic;

  Points0 = {
    {-0.00920594,  0.36459700,  0.0174376},  // fovea capitis
    { 0.03691960, -0.01011610, -0.0197803},  // anterior lateral condyle
    { 0.03001110, -0.00998133,  0.0186877},  // anterior medial condyle
    { 0.02009270,  0.34511400, -0.0387426},  // anterior greater trochanter point
    { 0.02783850,  0.18320400, -0.0217463},  // anterior shaft point
    {-0.02461770, -0.00623515, -0.0231383},  // posterior lateral condyle
    {-0.03211040, -0.00908290,  0.0246153},  // posterior medial condyle
    {-0.02643670,  0.35630800,  0.0014140},  // posterior head point
    { 0.01780310,  0.36194400,  0.0059740},  // anterior head point
    {-0.00197744,  0.38387300, -0.0031698},  // superior head point
    {-0.00316772,  0.34248600,  0.0114698},  // inferior head point
    {-0.02469710,  0.30335600, -0.0171113},  // medial lesser trochanter
    {-0.00969883,  0.34826800, -0.0462823},  // distal trochanteric fossa
    {-0.01959660,  0.36243100, -0.0441186},  // proximal posterior greater trochanter
    {-0.00084335,  0.32253400, -0.0641596},  // distal trochanteric fossa
    {-0.00431680,  0.35912600,  0.0036940}   // femoral COR
  PointNames = {

  Points1 = {
    { 0.2900, 0.4205, 0.0139},
    { 0.3220, 0.4332,-0.3786},
    { 0.2893, 0.4268,-0.3730},
    { 0.3599, 0.4429,-0.0050},
    { 0.3289, 0.4259,-0.1750},
    { 0.3062, 0.4872,-0.3703},
    { 0.2619, 0.4759,-0.3727},
    { 0.2900, 0.4405, 0.0139},
    { 0.3200, 0.4095, 0.0134},
    { 0.3100, 0.4295, 0.0314},
    { 0.2983, 0.4196,-0.0066},
    { 0.3089, 0.4599,-0.0355},
    { 0.3349, 0.4579, 0.0050},
    { 0.3329, 0.4679, 0.0175},
    { 0.3519, 0.4599,-0.0355},
    { 0.3075, 0.4235, 0.0139}
  BoundingBox =
    Type = BB_Cartesian;
    ScaleXYZ = {2, 2, 2};
    DivisionFactorXYZ = 5*{1, 1, 1};
  BoundingBoxOnOff = On;

This code constructs a transform, which deforms the source geometry into the target one using the thin-plate interpolation method and minimizes the distance between the selected key points (landmarks). This can be used when certain muscle attachment areas/points need to be scaled. Using this allows us improving the model by making some local features more accurate for the sensitive analyses. Please note that MyTransform object was included as a pre-transform as a rough scaling preceding the nonlinear RBF function, and it will be applied to the source entities, i.e. achieving the result of the previous step. Target bone is color-coded with the yellow color, initial linear scaling is grey, RBF-scaled bone is red. Tip: mouse-over in the Model View helps to see the name of the object.

However, it still possible to improve the fitting of the femur surfaces and, thus, improve the accuracy of the model. Looking at the Model View you can notice that the red and yellow surfaces are slightly different, e.g. at the femoral head region. This is caused by the nature of the interpolation and a low number of control points. The following section will describe how to utilize surface information for the construction of an improved scaling law.

Incorporating surface based nonlinearities into the scaling function

In this section, next improvement to the morphing is added by utilizing surface information. The surfaces will be requested to morphg into each other, which will at the same time deform all related soft tissue attachment points accordingly. The AnyFunTransform3DSTL class is used for this purpose. This class constructs an RBF transformation similarly to the AnyFunTransform3DRBF by using either corresponding vertices on the STL surfaces or seeding a number of vertices on one surface and finding a matching closest point on the second. For constructing a transformation using the vertices of STL surfaces, the surfaces have to be topologically equivalent, i.e. the surfaces have the same number of triangles and each neighbor and vertices represent the same features on both surfaces. For the latter option, we require an acceptable pre-registration transform, e.g. the RBF transform that was described previously, in order for the closest point search to make sense. Due to the implementation specifics most of the RBF recommendations apply to this class as well. More details about how to create this kind of transforms are described in appendix tutorial. However, for this example the recommended settings mentioned before will be used again.

Let us repeat the step from the previous section by adding one more surface to the visualization and another scaling step. You can download the model with all modifications here:

  AnyDrawSurf SurfaceMorphedSTL =
    FileName = "SourceFemur.stl";
    AnyFunTransform3D &ref = ..MySTLTransform;

    AnyFunTransform3DSTL MySTLTransform =
  PreTransforms = {&.MyRBFTransform};
  PolynomDegree = 1;
  RBFDef.Type = RBF_Triharmonic;
  AnyFixedRefFrame Input = {
    AnySurfSTL SourceSurf = {
      FileName = "SourceFemur.stl";
      ScaleXYZ = {1, 1, 1};
    AnySurfSTL TargetSurf = {
      FileName = "TargetFemur.stl";
      ScaleXYZ = {1, 1, 1};

  SurfaceObjects0 = {&Input.SourceSurf};
  SurfaceObjects1 = {&Input.TargetSurf};
  //FileName0 = "SourceFemur.stl";    // such definition was used previously
  //FileName1 = "TargetFemur.stl";    // such definition was used previously
  NumPoints = 1000;
  BoundingBox.ScaleXYZ = {2, 2, 2};
  BoundingBox.DivisionFactorXYZ = {1, 1, 1};
  BoundingBoxOnOff = On;

Please note again the transform from the previous section of this tutorial was included as a pre-transform, which means we will start working with the result of the previous step. Reloading the model, we can now see all steps of scaling in one place and can switch them on and off. For example, let us try to hide affine and RBF scaled femurs to see the final results:


If we just look at the green target surface and the blue STL-transformed surface, we can see that the surfaces now match each other very well. That means that now the subject-specificity will be taken into account in the inverse dynamics simulation. The final model can be downloaded here.

Finally, the only thing left is to include this scaling function into an actual model. Lesson 2 describes how this can be done.