Inaugural – Dissertation

In this thesis methods and applications of supervised learning for the segmentation and analysis of digital imagery coming from a variety of research domains are investigated. Segmentation and classification are important tasks in biomedical and industrial imaging and often provide the basis for recognition and quantification. Various specialized solutions exist for an enormous amount of distinct types of data and these are usually designed to meet and exploit given domain knowledge. In this work, an interactive supervised learning framework is proposed that is able to tackle multi-object segmentation and multi-class discrimination in a unified way. The method is general enough to cover a reasonable range of use cases in which local image descriptors are sufficient. The performance of the segmentation results is demonstrated on various data sets with distinct tasks to solve. This emphasizes the versatility of this approach to many biomedical and industrial data sets without requiring explicit image processing expertise and the need for custom programming. The approach builds upon a generic feature set that is able to characterize local cues such as color, texture and edges. To this end, an interactive tool that performs real-time processing on usual image sizes was developed, enabling domain experts to perform segmentation and classification tasks in an explorative fashion. No prior expertise in image processing is required since user interaction is facilitated via intuitive brush strokes. Once the algorithm/system has been trained, it can be applied to thousands of images with no further interaction with the user. The approach is limited to the segmentation of objects that can be discriminated based on local cues such as color or texture; but within this setting, the supervised framework yields surprisingly good results; on top of those, application-dependent post-analysis can be applied. The framework supports up to 4-dimensional multi-spectral data in an integrated fashion. In order to show the applicability and transferability of the method, several real world data sets – from very diverse imaging fields – are examined. Among them is the segmentation of tumor tissue from fluorescent wide-field microscopy, quantification of cell migration in confocal microscopy images for surveys on adult neurogenesis, segmentation of blood vessels in the retina of the eye, tracing of copper wires spread on tags for brand-owner authentication in an industrial context, and the application to image quality control for high-throughput siRNA screens. Furthermore an industrial problem is considered: a novel sequence classification procedure on the basis of localized frequency estimates is proposed for the process control and visualization of the sheet-feeding process for offset printing machines.