中国科学院自动化研究所机构知识库

理解大脑神经网络的结构和工作原理需要使用序列切片成像方法对脑组织样品进行三维重建。为了提高重建效果，序列脑组织切片需要被收取到硅片上进行后续的电镜成像。 目前以硅片为基底的脑组织切片收取主要通过人工收取方式，操作难度大且收取效率低。在此背景下，本文开展了脑组织切片自动收取系统的设计与相关实验研究，具有重要的学术价值与实际应用前景。主要研究内容包含：脑组织切片自动收取装置的机构设计、面向脑组织切片自动收取的显微图像目标检测方法与基于显微图像目标检测的脑组织切片自动收取系统设计。本文的主要内容和创新之处如下：

1. 脑组织切片自动收取装置的机构设计：目前已有的脑组织切片自动收取方案是以条带为基底的收取方式，该收取方式有后续操作繁琐、电镜成像质量不佳等问题。针对此问题，本文提出了新颖的基于圆形硅收取基底的脑组织切片自动收取方案，同时基于该收取方案研制了脑组织切片自动收取装置。所设计的脑组织切片自动收取装置的机械结构包括双挡板切片水槽、位置调节平台与硅片旋转机构。双挡板切片水槽用于与超薄切片机配合制备脑组织切片，同时限制脑组织切片条带在切片水槽的两个挡板间有序推进，位置调节平台用于将硅片旋转机构搭载的收取硅片以合适位姿置入切片水槽中收取脑组织切片。实验结果表明所设计的脑组织切片自动收取装置在开环收取状态下能够实现脑组织切片的有序收取。

 2. 面向脑组织切片自动收取的显微图像目标检测方法：为实现基于脑组织切片自动收取装置的闭环自动收取，首先需要开展面向脑组织切片自动收取的显微图像目标检测方法研究。针对脑组织切片的单类别目标检测任务与脑组织切片、切片水槽左挡板、右挡板的多类别目标检测任务，本文分别提出了基于深度学习的Simplified SSD目标检测算法与Form-invariant SSD目标检测算法。同时考虑到基于深度学习的目标检测算法需要大量的训练数据以达到较好的检测精度，而制备大规模脑组织切片显微图像数据集存在成本高、难度大等问题，因此本文采用通用数据增广方法与基于生成对抗网络的Cycle-GAN数据增广方法对所采集的脑组织切片显微图像数据集进行增广。实验结果表明所提出的基于数据增广方法的显微图像目标检测算法能实时且精准地检测显微镜视野中的各切片和挡板。

 3. 基于显微图像目标检测的自动收取系统设计：为实现基于脑组织切片自动收取装置的闭环自动收取并且提高切片收取的效率，本文基于显微视觉感知与反馈控制方法，为脑组织切片自动收取装置设计了配套的自动收取系统。该系统基于显微图像目标检测方法实时感知脑组织切片的制备与收取状态，基于该显微视觉感知结果，所设计的脑组织切片自动收取策略将自动调节收取装置到合适的工作状态以收取脑组织切片。实验证明在使用脑组织切片自动收取策略的情况下能使单片收取硅片收取的脑组织切片数量提升55%。电镜成像对比实验结果表明本文所采用的以硅片为基底的收取方式相比于以条带为基底的收取方法能取得更佳的电镜成像质量。

Understanding the structure and working principle of brain neural networks requires three-dimensional reconstruction of brain tissue samples using the array tomography method. In order to improve the reconstruction performance, the sequence of brain sections should be collected by silicon wafer for subsequent electron microscopic imaging. Currently, the collection of brain sections based on silicon substrate is mainly manual collection, which is difficult and inefficient to operate. Under this circumstance, this paper carries out the design and related experimental research of the automatic collection system for brain sections, which has important academic value and practical application prospect. And the main research contents include: the mechanical structure design of the automatic collection system for brain sections, the microscopic object detection method for the automatic collection of brain sections and the design of automatic collection system for brain sections based on the microscopic object detection method. The main contents and innovations of this paper are summarized as follows:

1. Mechanical structure design of the automatic collection device for brain sections: Currently, the existing automatic collection scheme of brain sections is based on tape substrate, which has the following problems: cumbersome follow-up operations, poor quality of electron microscope imaging and so on. In order to solve this problem, a novel automatic collection scheme of brain sections based on circular silicon substrate is proposed, and an automatic collection device of brain sections is developed based on this scheme. The mechanical structure of the proposed automatic collection device includes the knife boat with double baffles, the position adjusting platform and the silicon wafer rotating mechanism. The knife boat with double baffles is used to prepare brain section in cooperation with the ultra-microtome, while restricting the orderly advance of brain section between the two baffles of the knife boat. The position adjustment platform is used to place the silicon wafer loaded by the silicon wafer rotating mechanism into the knife boat to collect brain sections in a suitable position. The experimental results show that the designed collection device can achieve orderly collection of brain sections under open-loop collection.

2. Microscopic object detection method for the automatic collection of brain sections: In order to realize the closed-loop automatic collection of brain sections based on the designed collection device, it is first necessary to carry out research on microscopic object detection method for automatic collection of brain sections. For the single-category object detection task of brain section and the multi-category object detection task of brain section, left baffle and right baffle, this paper proposes Simplified SSD and Form-invariant SSD object detection algorithm, respectively. Considering that the deep learning based object detection algorithms require a large amount of training data to achieve better detection accuracy, and the preparation of large-scale microscopic image data-set of brain sections is difficult and expensive. So the generic data augmentation method and the Cycle-GAN data augmentation method based on the generative adversarial networks are used to augment the collected microscopic image data-set. The experimental results show that the proposed microscopic detection algorithm with data augmentation can accurately detect the brain sections and baffles in the microscope field in real time.

3. Design of automatic collection system for brain sections based on the microscopic object detection method: In order to realize the closed-loop automatic collection of brain sections based on the designed collection device, and improve the efficiency of section collection, this paper designs a matching automatic collection system based on micro-vision perception and feedback control methods. The automatic collection system senses the preparation and collection state of brain sections in real-time based on the microscopic object detection algorithm. And based on the microscopic perception results, the designed automatic collection strategy for brain sections could automatically adjust the collection device to the appropriate working state to collect brain sections. The experimental results show that the number of brain sections collected by a single silicon wafer can be increased by 55% when using the designed automatic collection strategy for brain sections. And the comparative experimental results of electron microscope imaging show that the silicon-substrate based collection method used in this paper can achieve better electron microscope imaging quality than the tape-substrate based collection method.