Mouse Brain Reconstruction and Analysis with Applications to Epilepsy Study

Biomedical images play an important role in biomedical research and diagnostics. However, the raw data acquired by imaging equipment sometimes are not suitable for direct observation and analysis. For example, the confocal microscope enables the observation of tissue in detail and reflects structures on the scale of a single cell or even finer. However, the raw data acquired by a confocal microscope usually contain multiple artifacts. These distortions include low SNR, irrelevant tissue clutter and geometric distortions. Image restoration and reconstruction algorithms, such as denoising, stitching and registration, are necessary before further analysis on these data. Another kind of widely used biomedical image is Magnetic resonance imaging (MRI). As an important non-invasive imaging technique, MRI facilitates the diagnosis of and research in diseases like brain cancer, Alzheimer’s and Parkinson’s. Reconstruction algorithms that turn the data in the frequency domain into the spatial domain are necessary after MRI scanning. Selecting proper parameters for these reconstruction algorithms is crucial to get high
quality MRI.
In this thesis, we focus on the image processing requirements on mouse brains for status epilepticus (SE) research. Epilepsy is a group of neurological disorders characterized by epileptic seizures. The rate of adverse outcomes of SE correlates with the duration of seizures, and thus early termination of SE is important. High-resolution 3D mouse brains provide details about SE development at multiple scales from cells, circuits, systems, to the whole brain level. Figuring out the pathways of SE development helps neuroscientists better understand the mechanism of SE and develop new drugs to terminate SE at an early stage.
Currently, the main way to investigate brain activity during SE at single neuron resolution is microscopy imaging. However, the penetration depth of some immunohistochemical neuron stains is limited to about 200 microns, and this requires mouse brains to be sliced before imaging. To better visualize and understand brain activity during SE, this thesis comprises three parts: 3D mouse brain reconstruction with microscopy data, auxiliary modality imaging to aid multi-brain analysis, and analysis of microscopy data. First, to recover the high resolution 3D mouse brain volumes, we propose tissue flattening and structure-based intensity propagation for 3D mouse brain reconstruction. Experiments are conducted on 367 multilayer sections from 20 mouse brains. The average reconstruction quality measured by the structure consistency index increases by 29% with the proposed structure-based intensity propagation. In order to better conduct multi-brain comparison and registration, an auxiliary imaging technique, MRI, is investigated in the second part. MRI is able to provide a complete 3D mouse brain volume before slicing. With the proposed parameter selection method, high quality synthetic MRI are reconstructed from measured data in the frequency domain. Finally, automatic cell detection enables neuroscientists to obtain cell activation information on the whole brain scale. To improve the detection accuracy in regions with densely-packed granule cells, we design a new center coding scheme for convolutional neural networks (CNN). With 3D mouse brain reconstruction and automatic cell detection, the 3D topology of cell activation is acquired, and this facilitates neuroscientists’ investigations of the mechanism of SE at multiple scales.