Deformable area based template matching with application to low contrast imagery

The exact positioning of patients during radiotherapy is essential for high precision treatment. Before each session, the patient must be accurately placed within the treatment device. The analysis of megavoltage X-ray images, portal images, can help to control this patient positioning. The particular problem of electronic portal imaging devices however, is that they provide imagery with extremely low contrast. Thus, common feature extraction schemes, such as simple edge detection, do not produce reliable results. Moreover, it is a very challenging task to design any robust feature extractor for such images, since a method must be found that reliably transforms the image information into a more useful form, whilst remaining robust to a large degree of image quality. To circumvent the step of feature extraction, the area-based method of least squares template matching (LSM) has been chosen. LSM is an iterative and area-based fitting method especially suitable for attaining very high precision, or for processing low-contrast, noisy, and blurred imagery. This thesis reviews the mathematical formalism of LSM and presents a framework for automatic quality control of the resulting match. This quality control—a component often missing in commonly used image matching methods—is achieved by self-diagnostic measures supervising the iterative procedure. The application of LSM to the problem of patient positioning was thoroughly investigated. Three issues had to be tackled: the calibration of the newly acquired portal image; the checking of the field shape; and the displacement measurement of bony structures. Precise calibration has been achieved by matching the field edge region to a reference shape, where LSM has proven to yield robust and accurate results. Hence, it was feasible to check the field shape by simply computing the normalized cross correlation between the actual and the prescribed field. In several test series, the suitability of LSM was assessed to measure patient displacement. The area-based matching of reference structures to the treatment image was carried out using an affine or congruent transformation model. A very promising success rate of over 90 % was achieved using clinical test data which consisted of roughly 60 image series with a total of 500 portal images. The optimum sources for reference data are the actual treatment settings for the calibration, the planning data for checking the field shape, and digitally reconstructed radiographs (DRR) for the displacement measurements. Due to missing system integration however, an interim approach in terms of employed reference data had to be chosen. …