UC Irvine

Cerebral microbleeds characterize a heterogenous population of diseases responsible for serious brain injury and death. The current state of the field includes clarifying the origin and progression of microbleed associated diseases through identification of disease specific biomarkers and vascular contributors to hemorrhagic activity. Microbleeds have been associated with dysfunction of the microvasculature in several studies. While some vascular information can be gathered clinically (e.g. through magnetic resonance imaging, MRI), identification of capillary level information requires the use of ex-vivo analysis. While the current standard is 2D histology of thin (e.g. 40 micron) brain sections, progression of research in this field may benefit from a tool to represent microbleeds and surrounding vascular structure systemically in the whole, unaltered organ.

To explore brain architecture and pathology, a consistent and reliable methodology to visualize the three-dimensional cerebral microvasculature is beneficial. Perfusion-based vascular labeling is quick and easily deliverable. However, the quality of vascular labeling can vary with perfusion-based labels due to aggregate formation, leakage, rapid photobleaching, and incomplete perfusion. Here, we describe a simple, two-day protocol with perfusion-based labeling that facilitates whole-brain, three-dimensional microvascular imaging and characterization. The combination of retro-orbital injection of Lectin-Dylight-649 to label the vasculature, the clearing process of a modified iDISCO+ protocol, and light-sheet imaging, collectively enables a comprehensive view of the cerebrovasculature. We observed an ~18-fold increase in contrast-to-background ratio of Lectin-Dylight-649 vascular labeling over endogenous GFP fluorescence from a transgenic mouse model. With light-sheet microscopy, we demonstrate sharp visualization of cerebral microvasculature throughout the intact mouse brain. Our tissue preparation protocol requires fairly routine processing steps and is compatible with multiple types of optical microscopy.

Further, machine learning techniques provide an advantage to large data sets which may apply to the delineation of vascular structures. To test this theory, we use deep learning algorithms on vascular images taken by our lab of cleared whole organs labeled with Lectin Dylight and imaged with lightsheet microscopy. Convolutional neural networks (CNNs) via transfer learning were used for the multi-class classification of mouse organs from vasculature images. The datasets were put through multiple CNNs pre-trained on ImageNet including GoogLeNet, Inceptionv3, and NASNetLarge. Among them, GoogLeNet was the best because of the high accuracy among training and test datasets and relatively quick prediction rate.