CRCNS Research Proposal: A Unified Framework for Unsupervised Sparse-to-dense Brain Image Generation and Neural Circuit Reconstruction

要了解信息是如何在大脑的神经回路中处理和传播的，就需要精确地绘制出大量密集、混杂的神经元之间的连接。尽管绘制大脑回路的努力已经持续了一个世纪，但对这些复杂网络的解剖仍然具有技术挑战性和耗时性。该项目建立在遗传，成像和计算方法的最新进展的基础上，开发了一个统一的框架，用于从3D图像数据中以单神经元分辨率可靠地重建遗传识别的神经元及其连接。由此产生的计算方法将适用于生物和生物医学图像分析中的广泛问题。该项目的技术目标分为四个方面。第一个推力产生了大量的超分辨率三维（3D）图像的整个果蝇大脑，在每个连接的神经元标记的多光谱trans-Tango/Bitbow标记。第二个重点是创建计算机视觉和无监督机器学习算法来生成神经元跟踪，以及帮助高效的人类校对和纠错的软件。这将允许从相对稀疏的Brainbow标记生成金标准神经元追踪和分割作为训练输入。第三个推力创建了一个生成式机器学习模型，以创建逼真的Bitbow神经元，并基于稀疏注释训练输入合成具有各种标记密度的图像。第四个重点是开发一个注释机器学习模型，以机器学习生成的Bitbow图像及其相应的地面实况注释作为训练输入，重建密集标记的trans-Tango/Bitbow果蝇大脑。这些努力共同形成了一个新的计算框架，能够准确地自动重建密集标记的神经回路。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Understanding how information is processed and propagated in neural circuits in the brain, requires the connections among extremely large numbers of densely packed, intermingled neurons to be accurately mapped out. Despite the century-long effort to map brain circuits, dissection of these complex networks remains technically challenging and time-consuming. This project builds on recent advances in genetic, imaging, and computational methods to develop a unified framework for reliable reconstruction of genetically identified neurons and their connections at single-neuron resolution from 3D image data. The resulting computational methods will be applicable to a wide range of problems in biological and biomedical image analysis.The technical aims of the project are divided into four thrusts. The first thrust generates a large number of super-resolution three dimensional (3D) images of whole Drosophila brains, in each of which connected neurons are labeled by multispectral trans-Tango/Bitbow labeling. The second thrust creates computer vision and unsupervised machine learning algorithms to generate neuronal tracing, and software to aid efficient human proofreading and error correction. This will allow the generation of gold-standard neuronal tracing and segmentation from relatively sparse Brainbow labeling as training inputs. The third thrust creates a generative machine learning model to create realistic Bitbow neurons and synthesize images with various labeling densities based on the sparse annotation training inputs. The fourth thrust develops an annotation machine learning model to reconstruct the densely labeled trans-Tango/Bitbow Drosophila brains with the machine-learning-generated Bitbow images and their corresponding ground truth annotations as training inputs. Together, these efforts form a novel computational framework to enable accurate automatic reconstruction of densely labeled neural circuits.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.