毕业生知识库

在全脑尺度上解析神经连接图谱是揭示脑工作原理的重要基础，有利于推动新一代智能产业的发展。使用电子显微镜进行成像的脑微观重建是在纳米尺度上还原神经三维连接的有效方法。目前，随着脑科学和显微成像技术的不断发展，脑微观重建的体量越来越大，脑组织序列切片的数目越来越多。对收集到的大量脑组织序列切片进行快速准确检测是后续进行电子显微镜自动拍摄的重要前提，人工检测会耗费大量人力和时间。本文提出了基于深度学习的脑组织序列切片自动检测算法，能够准确地对序列切片进行检测，为后续电子显微镜自动拍摄提供精确的切片位置，加速自动拍摄过程，同时减少人工参与，节省人力和时间。

本课题作为脑微观重建中自动采集流程的一部分，取得的主要成果包括：1）针对褶皱、破损等异常切片和正常切片因形态相似而不易区分，以及切片之间互相粘连导致边界不清而难以分辨等问题，提出了多频通道注意力机制，通过离散余弦变换得到特征图各个通道不同的频域分量作为通道注意力来加强特征图中的重要信息，增强了网络对切片的识别能力。在自制数据集上的实验结果表明，该方法能够对序列切片实现更加精确的检测。2）针对神经网络在面对新的脑组织序列切片检测任务时，需要重新制作数据集训练网络的情况，提出了一个主动学习算法，通过在原有的网络上添加损失预测模块预测的网络损失以及类别后验概率加权的实例分割掩码不确定度作为样本的价值量，从而只需要标注少量价值量大的样本制作数据集便可使网络达到预期的性能。实验表明，通过该方法只需30%的样本作为数据集便可使网络达到预期性能。3）针对脑组织序列切片自动检测算法在自动采集流程中的应用问题，提出了一套软件系统，能够调用检测算法对序列切片进行检测并将结果输出为后续采集可以直接使用的特定格式，从而将检测算法嵌入到了脑微观重建的自动采集流程中。

The analysis of neural connection maps on the whole brain scale is an important basis for revealing the working principle of the brain, which is conducive to promoting the development of a new generation of intelligent industry. Microscale reconstruction of the brain using electron microscopy is an effective method to restore neural three-dimensional connections at the nanoscale. At present, with the continuous development of brain science and microscopic imaging technology, the volume of brain microscale reconstruction becomes bigger and the number of brain tissue serial sections is much more. Rapid and accurate detection of a large number of collected brain tissue serial sections is an important basis of automatic electron microscopic imaging process. Manually detecting these serial sections takes a lot of labor and time. This paper proposes an automatic detection algorithm based on deep learning, which can accurately detect serial sections, provide accurate section positions for automatic electron microscope imaging, speed up the automatic imaging process, reduce manual participation, save labor and time.

As a part of the automatic electron microscopic imaging process in brain microscale reconstruction, the main achievements of this paper include: 1)To solve the problems that abnormal sections such as wrinkles and damage are difficult to distinguish from normal sections due to their similar shapes, and the boundaries between sections are not clear and difficult to distinguish due to the adhesion between sections, a multi-frequency channel attention mechanism is proposed. The frequency domain components of different channels are used as channel attention to enhance the important information in the feature map and enhance the networks ability to recognize sections. The experimental results show that the method can achieve more accurate detection of sequence sections. 2)An active learning algorithm is proposed for the situation that the neural network needs to be fine-tuned using new dataset when facing the new brain tissue serial sections detection task. A loss prediction module is added to the instance segmentation network to predict the loss. The loss predicted is added to the category posteriori probability weighted instance segmentation mask uncertainty as the value of a sample. Selecting the samples with large value to label constitutes the dataset, which enables network to have the expected performanc. Experimental results demonstrate that this method can achieve the desired performance of the network with only 30% samples as dataset. 3)For brain tissue sections automatic detection algorithm in the application of automatic data collection process, a software system is proposed. The system uses detection algorithm to detect serial sections and result can be used directly for subsequent acquisition of a particular format. Thus the detection algorithm is embedded in the automatic electron microscopic imaging process in brain microscale reconstruction.