ETH Zürich, 2011. — 154 p.In this thesis we investigate the whole image processing pipeline for neuronal geometry extraction and synapse detection in electron microscopy images. Advancements in automated sample preparation and image acquisition for electron microscopy enable recording of large data sets. This process is especially important for the field of computational neuroanatomy and connectomics, as the analysis of neuronal connections requires imaging of large volumes with a resolution sufficient for synapse detection. Manual processing of electron microscopy data is time-consuming and becoming the main bottleneck in gaining new insights into the functional structure of the brain. Automated processing of biological electron microscopy images is challenging due to the rich texture, low signal to noise ratio and the great variability of image characteristics depending on sample preparation and animal type. To enable quantitative evaluation of the data, the images are corrected against lens distortions, stitched, and aligned. Structures of interest are then segmented and grouped across serial sections to extract the 3d geometry. The proposed registration methods employ unsupervised approaches to identify artifact signals like non-linear distortions, cracks, or staining blurs. We demonstrate that identification of these signals leads to superior registration results compared to state-of-the-art methods. The distortion correction enables structure preserving mosaicing with sub-pixel precision. The non-linear distortion field is estimated from overlapping image areas and does not require special calibration samples.
Чтобы скачать этот файл зарегистрируйтесь и/или войдите на сайт используя форму сверху.