Cellular/Vascular Reconstruction Using a Deep CNN for Semantic Image Preprocessing and Explicit Segmentation

Abstract

Maps of brain microarchitecture are important for understanding neurological function, behavior, and changes due to chronic conditions, such as neurodegenerative diseases. New high-throughput microscopy techniques produce whole organ data sets imaged at sub-cellular resolution. The resulting volumetric data is composed of densely packed cells and interconnected microvascular networks. The data size and complexity makes manual annotation impractical and automatic segmentation challenging. In this paper, we propose a processing pipeline to segment, reconstruct, and analyze cellular and vascular microstructures in large rodent brain volumes. We first introduce a fully-convolutional, deep, and densely-connected encoder-decoder for pixel-wise semantic segmentation as a pre-processing step in our pipeline. Excessive memory complexity is mitigated by compressing the features passed through skip connections, resulting in fewer parameters and enabling a significant performance increase over prior architectures. We then quantify the pipeline’s scalability, accuracy, and reliability for extracting explicit cellular and vascular images, including vessel connectivity. Finally, we demonstrate the viability of this processing pipeline on a large (1 $\mathrm {mm}^{3}$) region of the mouse somatosensory cortex as a proof of efficacy.