The knee alignment angle, defined by the mechanical axes of the femur and the tibia, is of high importance in orthopedics and traumatology, in particular for pre-operative planning and post-operative follow-up assessment. It serves as a predictor for pre-arthritis and the post-operative angulation is known to be associated with the clinical outcome.
So far, only manual and semi-automated methods of measuring the alignment angle exist. The most serious drawback of these techniques is that they lack reproducibility. The points in the anatomical structure, which define the angle to be measured, are only vaguely defined and cannot clearly and precisely be identified. The resulting variability between repeated measurements precludes the detection of small changes. A fully automatic measurement method of axis alignment that provides a consistent definition of anatomical landmarks would eliminate inter and intra reader variabilities caused by human interpretation and lead to more accurate and reproducible results.
In the course of this thesis, a novel method for the automatic measurement of alignment angles is developed and prospectively tested. It allows a fully automatic assessment of knee alignment angles in full-limb radiographs with high precision.
The positions of the lower limb bones and joints are estimated using Sparse Markov Random Field Appearance Models. They are able to detect anatomical structures by configurations of interest points, taking their spatial arrangement and local appearance into account.
Based on the coarse position estimates of the bones, their contours are delineated by Active Shape Models, controlled by ongoing estimates of the reliability of the model. The regions around the joints are refined using submodels.
Landmarks are identified by their index and can be matched between different instances of a shape. Hence, the defining points of the axes can be located in a straightforward and repeatable way, when they are directly represented by landmarks on the contour and annotated manually in the training phase on any instance of the bone.
Overlapping structures and the compound nature of the large radiography acquisition which results in partially missing data and changing intensities let standard ASMs fail. For this reason, a search procedure is introduced, which is controlled by ongoing estimates of the fit confidence during the search, leading to an improved result robustness, even if the spatial initialization is poor and the structures of interest are partially cropped or occluded.
Experimental results show that the automatic assessment of the knee alignment angle allows for an accurate and observer independent quantification with high precision and improves the detection of small changes.