Automatic Modeling of Anatomical Variability for Object Localization in Medical Images

Object localization in medical images plays an ever increasing role as prerequisite for many automatic image processing routines like segmentations or automatic measurements. Yet, due to the strong challenges imposed on this task through different imaging modalities likes X-ray, CT, or MRI, and the anatomical and pathological differences between different patients, it is often still performed manually. To provide a general and automatic method for object localization which is capable of dealing with the above stated challenges, this work proposes an automatic generation and discriminative training of models for object localization with the generalized Hough transform. This discriminative generalized Hough transform is able to capture a large amount of the object's variability and to train robust and precise localizations models for bone as well as soft tissue targets in 2D or 3D images. The method is tested on various targets and modalities, amongst others the knee and clavicle in X-ray, the vertebrae in C-arm CT, and the femoral bone in MRI. In addition, a comparison with a state-of-the art method is performed and the procedure is benchmarked on a public database.