HipOP: A Framework for Automatic Classification of Bone Density Based on 2D X-Ray Images
The risk of elderly patients to suffer hip or femur fractures is over average as well as the frequency of decreased density and stability of their bones due to osteoporosis. Therefore therapy planning has to consider the stability of a possible implant footing to avoid complications caused by relaxed implants. For this task reliable and quantitative knowledge of the individual status of the bone tissue is crucial. Although two methods exist to gather this knowledge, the dual energy x-ray absorptiometry (DXA or DEXA) and the CT bone densitometry, these methods are expensive and not always available in clinical routine. The conventional x-ray image therefore remains the typical and common routine method for diagnosis and therapy planning.
Hip overview x-ray images are the routine diagnostic method in the diagnostic process of hip fractures and therefore always available at no extra cost.
The core challenge of this project is the reliable mapping of regions of measurement onto the femur. The definition of the areas to measure bone density is based on the findings of Manmohan Singh et. al.  concerning the changes in trabecular pattern of the upper end of the femur.
Detecting the femur in an x-ray image is a difficult problem to solve in terms of an exact segmentation step but can be reduced to a registration problem since the reproducible mapping of the measurement pattern does not need a definite segmentation but a good global positioning of a femur template.
In our approach a mutual information based registration  algorithm uses a template image to detect the femur in the x-ray image. It is optimized regarding the gray value distribution in the template image as well as the transformation.
Having the transformation of the template to the x-ray image and knowing the positions of the regions of measurement in the template makes it possible to map these regions onto the x-ray image.
As x-ray images are 2D projections of a 3D body, it is not possible to calculate with absolute gray values. The estimation of bone density values is based on relations between different regions where the pixel values within one region are averaged. The measurement of regions is necessary because noise can alter single pixels extremely.
- Univ.-Klinik für Unfallchirurgie und Sporttraumatologie Landeskrankenhaus – Universitätskliniken – Innsbruck Vorstand: Univ.-Prof. Dr. Michael Blauth
- AO Foundation - Davos/Switzerland
State : Feasibility study finished. Main project started.
Funded by: FFF/HITT