Abstract
In the underlying dissertation a new approach towards model-based image processing of high-resolution CT and MRI images is described. The aim of this project was to develop a semi-automatic segmentation approach of medical data that allows the reconstruction of an irradiation model for the proton therapy of eye cancer. The usage of the image-processing algorithm was to require as few as possible interactions by a user while the correctness and the usability were to be optimized.
Especially, the precision requirements were very high in order to leverage the high dose precision of proton therapy. On the other side, a viable alternative to the manual segmentation was to be found as the workload connected to the manual evaluation of 150 to 300 image slices were unacceptable.
The investigations had shown that a critical factor to achieve the aimed at targets was the preliminary knowledge about the image material and the organs that are to be segmented. Based on the case-based reasoning approach a new image processing technique has been developed that performs a comparative segmentation using previously segmented data sets. These data sets are similar to the data set that is to be segmented with respect to content and imaging characteristics. They had been segmented by an expert within the clinical routine and are now reused as exemplary expected results. For the segmentation of a current data set, one of the exemplary solutions is retrieved from the case database using similarity criteria. Image type, slice orientation, and similarity in eye proportions were used to calculate the similarity of the data sets. To perform the segmentation a modified snake approach was adopted. Internal and external energy definitions were based on comparative similarity measures to perform a comparative segmentation that takes the expected segmentation result into account.
The usage of three-dimensional exemplary segmentations allows the inclusion of comprehensive knowledge about the expected properties of the image data, the representation of the organs of interest, the characteristics of the boundaries, the spatial arrangement, and the geometry of the organs. The case-based segmentation approach proved to be comfortable and user-friendly. No additional work was necessary to generate new cases for the case database after a data set had been segmented and an expert had confirmed the correctness of the segmentation. Adjustment to other organs and tasks can be achieved by replacing the according case database.
The tests proved the model-based segmentation approach to be precise and reliable. Only intra-individual manual segmentations provided a higher consistency and correctness. Still, the model-based segmentation approach achieved better results than inter-individual manual segmentations. The segmentation approach had been extended by a model-based interpolation and registration reusing major parts of the methodology to achieve multi-modal image integration. The performed tests confirmed the good results of the described model-based image processing approach as a reliable technique for the segmentation, interpolation, and registration of medical data sets. |
