Jan Hendrik Moltz

• This PhD thesis makes contributions in the field of medical image analysis for assisting CT-based staging and follow-up examinations of patients under chemotherapy. First, an algorithm for semi-automatic segmentation of liver metastases in CT images is presented. The runtime was kept within a clinically acceptable limit of less than 1 s on average. The method is able to deal with inhomogeneous density distributions and prevents leakage through the liver boundary. A comprehensive evaluation was performed on 371 lesions with manual segmentations. The method produces results of similar quality as other state-of-the-art methods but is significantly faster. In order to accelerate follow-up examinations in the clinic, I implemented a framework for automatic lesion tracking. For a segmented baseline lesion, it identifies the corresponding lesion in follow-up, automatically initializes the segmentation, and performs a plausibility check. No other framework automatizes follow-up examinations to this degree. A problem analysis, examining the change of 994 lesions under chemotherapy, and a simulation of similarity measures on a lesion phantom motivate the presentation of a template matching algorithm tailored to this problem. The method is validated from a technical point of view on 207 cases before reporting a user study that evaluated possible benefits for the clinical workflow. For validating segmentation algorithms, manual delineations are often used, but their high variability makes it difficult to achieve reliable statements. To quantify this problem and to overcome it in practice, I present a generalization of the MICCAI score that takes the variability of multiple reference segmentations into account for each case. An analysis of the variability in manual delineations of ten experts is performed. The thesis is concluded by a concept and a validation study for a tool that allows experts to generate probabilistic reference segmentations.