One aim of this project is to analyze a large set of medical image data with respect to the anatomy of the knee joint. The aim is to determine the variation in shape of the knee and knee joint space, respectively bone and cartilage, between distal femur and proximal tibia. Clusters of similar shapes have to be determined in order to design a limited set of knee spacers that fit a wide range of the osteoarthritic population.

Data selection

In order to detect different clusters of similar shapes at least 500 MRI datasets need to be processed. The datasets are taken from The Osteoarthritis Initiative (OAI) database. The OAI database contains about 5000 patients. Therefore, a selection has been made based on the Kellgren-Lawrence OA-Score which is available for almost all patients. The Kellgren-Lawrence score differentiates between five grades. For each of the five grades 138 female and 145 male patients have been randomly selected resulting in a preselection of 1415 right knee MRI datasets. In a first step 500 of these datasets are processed and analysed. As MRI protocol SAG_3D_DESS_WE (sagittal 3D dual-echo steady state with selective water excitation) is used for bone and cartilage segmentation.

Data processing

Bone and cartilage of distal femur and proximal tibia are segmented automatically using Statistical Shape Models [1]. Errors in the automatic segmentation are corrected manually. Additionally, for each knee landmarks of the insertion sites of anterior cruciate ligament (ACL) and posterior cruciate ligament (PCL) are placed by hand. A schematic overview is given in figure 1.

Fig. 1: Schematic overview of data processing

The following pictures illustrate the single steps of the pipeline.


Fig. 2: Bone segmentation after automatic segmentation (left) and after manual postprocessing (right)


Fig. 3: Cartilage segmentation after automatic segmentation (left) and after manual postprocessing (right)


Fig. 4: Insertion sites of ACL and PCL (left). Result of data processing (right). Cartilage thickness is displayed in red. ACL and PCL insertion sites are replaced by a representative point.



The aim is to find geometrical clusters in (very) high dimensional data that complicates meaningful clustering. Therefore, a principal component analysis (PCA) is done to reduce the dimensionality. But the PCA is a global technique. Every single point has the same influence on the result. Hence, the idea is to restrict the geometry to a region of interest.

Nevertheless, the results still have too many dimensions for clustering. For that reason, correlation and regression analysis between geometry and clinical parameters are done to achieve further reduction of dimensionality.





Fig. 5: Analysis is restricted to tibial and femoral parts of the respective condyle compartment.



[1] Seim, H., Kainmueller, D., Lamecker, H., Bindernagel, M., Malinowski, J., & Zachow, S. (2010). Model-based auto-segmentation of knee bones and cartilage in MRI data. Proc. MICCAI Workshop Medical Image Analysis for the Clinic - A Grand Challange, 215-223.