Model Based Autosegmentation of Brain Structures in the Honeybee Apis Mellifera


The honey bee standard brain (HBS) serves as a 3D platform for the visualisation of neuronal morphologies of different brain preparations. The registration of single neurons into the HBS requires a time-consuming manual segmentation of neuropil boundaries in confocal image data. Therefore it would be desirable to automate the segmentation process. A model-based auto-segmentation approach was succesfully applied for medical data by Lamecker et. al [2]. This method uses a-priori information about the shape of the segmented structure and features of the image data during the segmentation process. In this work, I want to figure out, if this method can be adapted to confocal image data. Therefore I establish it for the segmentation of a single bee brain structure, the right median calyx. At first I was able to solve the problem of the initialization of the model-based segmentation by an affine registration of the individual confocal dataset to the standard grey value dataset. In order to enlarge the variability of an existing statistical shape model of the calyx consisting of 18 training shapes, I extended it by another 27 training shapes. For the adjustment of this model I analysed grey value profiles along surface normals of the training data. On the basis of a profile model I developed a displacement algorithm for the adjustment of the model to a confocal image dataset. Then I tested the automatic segmentation with the different sized shape models and the developed displacement algorithm for 45 confocal image datasets. In 80% of the segmentations, the result was good compared to the manual segmentations. In some cases the profile model failed. Furthermore there were sometimes large maximal distances between manual and automatic segmentation, because the profile model did not fit the particular profile. The segmentation result could be improved with the extended model.

Bachelor thesis in Bioinformatics, Freie Universität Berlin