Large-Scale Segmentation of DNA-Labeled Brain Microstructures Using Deep Semantic Preprocessing

Mapping brain architecture is important for understanding neurological function and behavior, particularly when studying neurodegenerative diseases. High throughput microscopy, such as \textit{knife-edge scanning microscopy} (KESM) and \textit{milling with ultraviolet excitation} (MUVE), provides the potential for whole organ imaging at microscopic resolution. The size and complexity of the resulting images makes manual annotation impractical and automatic segmentation challenging. Densely-packed cells with heterogeneous sizes and shapes, combined with complex interconnected vascular networks, pose a problem for current localization and segmentation algorithms. In addition, large teravoxel data sizes necessitate the development of fast algorithms that are largely unsupervised. In this dissertation, a fast and robust cell and microvascular segmentation framework is implemented to quantify tissue microstructure in large microscopy images. The proposed method is validated using thionine-stained rat brain tissue.

Deep fully-convolutional neural networks are used as a pre-processing step to enhance the contrast of cellular and vascular components in the thionine-stained data. In this dissertation, a deep and densely-connected fully-convolutional encoder-decoder is designed for pixelwise classification. The trained network reliably distinguishes between cellular and vascular structures. Then a GPU-based cell localization algorithm is designed to identify the three-dimensional positions of cells. It is demonstrated that the proposed ``deep preprocessing'' framework significantly improves the accuracy of relative to the brain microvasculature.