The ability of accurate localize and characterize cells in light sheet fluorescence microscopy (LSFM) image is indispensable for shedding new light on the understanding of three dimensional structures of the whole brain. In our previous work, we have successfully developed a 2D nuclear segmentation method for the nuclear cleared microscopy images using deep learning techniques. Although the convolutional neural networks show promise in segmenting cells in LSFM images, our previous work is confined in 2D segmentation scenario and suffers from the limited number of annotated data. In this project, we aim to develop a high throughput 3D cell segmentation engine, with the focus on improving the segmentation accuracy and generality. First, we will develop a cloud based semi-automatic annotation platform using the strength of virtual reality (VR) and crowd sourcing. The user-friendly annotation environment and stereoscopic view in VR can significantly improve the efficiency of manual annotation. We design a semi-automatic annotation workflow to largely reduce human intervention, and thus improve both the accuracy and the replicability of annotation across different users. Enlightened by the spirit of citizen science, we will extend the annotation software into a crowd sourcing platform which allows us to obtain a massive number of manual annotations in short time. Second, we will develop a fully 3D cell segmentation engine using 3D convolutional neural networks trained with the 3D annotated samples. Since it is often difficult to acquire isotropic LSFM images, we will further develop a super resolution method to impute a high resolution image to facilitate the 3D cell segmentation. Third, we will develop a transfer learning framework to make our 3D cell segmentation engine general enough to the application of novel LSFM data which might have significant gap of image appearance due to different imaging setup or clearing/staining protocol. This general framework will allow us to rapidly develop a specific cell segmentation solution for new LSFM data with very few or even no manual annotations, by transferring the existing 3D segmentation engine that has been trained with a sufficient number of annotated samples. Fourth, we will apply our computational tools to several pilot neuroscience studies: (1) Investigating how topoisomerase I (one of the autism linked transcriptional regulators) regulates brain structure, and (2) Investigating genetic influence on cell types in the developing human brain by quantifying the number of progenitor cells in fetal cortical tissue. Successful carrying out our project will have wide-reaching impact in neuroscience community in visualizing and analyzing complete cellular resolution maps of individual cell types within healthy and disease brain. The improved cell segmentation engine in 3D allows scientists from all over the world to share and process each other?s data accurately and efficiently, thus increasing reproducibility and power.
This proposal aims to develop a next generation cell segmentation engine for the whole brain tissue cleared images. Our proposed work is built upon our previous 2D nuclear segmentation project using deep learning techniques. However, we found that our current computational tool is limited in 2D segmentation scenario and insufficient of annotated training samples. To address these limitations, we will first develop a cloud-based semi-automatic annotation tool with the capacity of virtual reality. Our annotation tool is designed to be cross- platform, which allows us to partner with ?SciStarter? (the largest citizen science projects in the world) and acquire large amount of cell annotations from the science enthusiastic volunteers. Meanwhile, we will develop next generation 3D cell segmentation engine using an end-to-end fully connected convolution neural network. To facilitate 3D cell segmentation, we will also develop a super resolution method to impute an isotropic high- resolution image from a low-resolution microscopy image. After the development of 3D cell segmentation engine, we will continue to improve its generality by developing a transfer learning framework which enables us to rapidly deploy our 3D cell segmentation engine to the novel microscopy images without the time-consuming manual annotation step. Finally, we will apply our segmentation tool to visualize and quantify brain structure differences within genetically characterized mouse and human brain tissue at UNC neuroscience center. In the end of this project, we will release the software (both binary program and source code) and the 3D cell annotations, in order to facilitate the similar neuroscience studies in other institutes. Considering the importance of high throughput computational tools in quantifying three dimensional brain structure, this cutting- edge technique will be very useful in neuroscience research community.