Individualized reconstruction

		Automatic bone segmentation on a sample slice and corresponding 3D models Bone and skin models of the full leg Topological optimisation of simplex meshes Spline-based definition of muscle attachment areas Hip bones tracked position on four dynamic slices at a sample position

For both the diagnosis and the surgical planning, an accurate estimate of hip anatomy and motion are required. Orthopaedists can use animated 3D models, prior to joints surgeries, to evaluate their task and generally to reduce the overall time of the surgical operation. The long-term objective is to model, analyse and visualise human joint motion in-vivo and non-invasively. This work aims at investigating methods for the anatomical modelling of the different components of the musculoskeletal system (mainly bones, muscles, fat and skin) and for the kinematical analysis of the rigid and soft active motion of these components, from medical images.

Anatomical modelling of the musculoskeletal system from static MRI

We focus on discrete deformable models-based techniques allowing a good shape and motion control for large deformations, as well as light computational expenses for simulating complex and interrelated models. Generic models are built interactively and then used as discrete deformable models to reconstruct any individual models automatically. We developed a first technique for automatic bone segmentation from MRI: after a coarse landmark-based initialisation and Bookstein's Thin-Plate-Splines interpolation, the generic model is iteratively deformed using deformation spheres with decreasing radius. The deformation is driven by the external energy of the model, measuring the matching between the model and features in the MRI volume, based on oriented gradient images and model normals. We compared manual and automatic segmentation on four different generic pelvises and femurs by doing cross tests. The average difference is less than 15% of the total number of voxels.

We are currently extending the framework for the modelling of all tissues, using multiresolution simplex meshes, firstly described by Delingette et al. These models allow a global-to-local, thus more robust, registration of generic models to individual images. Our contribution includes:

• The implementation of implicit methods for the resolution of motion equation
  
• The definition of external forces based on intensity profile registration.
  
• The definition of internal forces such as volume conservation, shape memory and smoothing forces.
  
• The development of methods for the optimisation of mesh topology
  
• The implementation of collision detection and response schemes suited for simplex meshes.
  
• The use of multiresolution in the deformation and collision handling processes.
  
• The definition of attachments areas between models (i.e. muscle attachments) based on splines.
  

Kinematical analysis of the musculoskeletal system from dynamic MRI

We have developed an image-based method to track hip bones motion. Our method is based on the combination of temporal information in dynamic MRI and spatial information in static MRI. Static MRI is a prior knowledge on the spatial configuration of organs. As bones motion is rigid, there is a spatial transformation (six parameters) between dynamic MRI slices and the static MRI volume that registers voxels of a given bone. We have tested different similarity metrics that can be applied to the intensity images or to the gradient vector maps. The amoeba optimisation method is used to compute automatically solution parameters that minimise the similarity metric inside a mask where the transformation is considered to be rigid. Our method was implemented with a multi-resolution approach. We have done a clinical study with ethics approval on six volunteers in order to validate the method (in terms of accuracy and robustness) and to optimise the acquisition (number of planes, resolution) and the tracking (similarity metric). Bones relative positions tracked in real-time MRI were compared to the gold-standard positions measured in sequential MRI for a fixed volunteer posture. We obtained a final error of 3deg in terms of relative position between femur and pelvis . The final dynamic protocol was a fast gradient echo sequence with balanced gradients (bFFE, TR/TE 3.5/1.1ms, Flip angle 80deg, pixel size 4.7 x 2.6mm, partial Fourier reduction factor of 0.65 in read direction, SENSE acceleration factor of 2, frame rate = 6.7 frames/s).