Zusammenfassung: | Conventional light microscopy is commonly used and one of the most powerful techniques for the inspection and diagnosis of biological samples. However, it is constrained by several factors, among which: only a small part of a sample is visualized at a time, magnifications are limited by the available microscope objectives, the slide is irreproducible and needs to be physically present for diagnosis. One goal of virtual microscopy is to solve these issues by replacing the direct work with the microscope and the slide with the use of digitized probes. For this purpose, the slide is scanned once by a fully automated microscope and saved digitally. The scanning process of the slide is done using computer-controlled stages. Depending upon the magnification, more than ten thousand images are thus acquired. To generate the virtual slide, these fields of view have to be composed into a large mosaic. However, motorized stages have positioning errors that can add up during the scanning process, leading to offsets in the final virtual slide. These offsets disturb the diagnosis and can even render it erroneous. This research deals with such impediments in virtualization. Notable is that registration in virtual microscopy poses quite a different problem compared to common registration or mosaicing problems found in panoramic image stitching or general medical registration problems from computed tomography for example. The emphasis of this research is on the correct alignment and positioning of all fields of view. Contrary to state of the art algorithms, in this work the alignment is not approached sequentially but as an optimization problem that incorporates the whole slide into a single mathematical model. This model and thus the positioning of every single image is solved by applying error minimization techniques in one single step. Furthermore, a new physically exact transformation method for the registration of differently resolved and magnified images is presented. This method relies upon an exact analysis of sampling frequency and bandwidth, leading to an integrated determination approach of transformation parameters. For the evaluation a novel method that builds upon the generation of synthetic slides is presented. By applying this method, ground truth positioning information is retrieved that is used to compare different mosaicing schemes. |
---|