毕业生知识库

      作为一种关键的纳米级成像技术，序列切片电子显微镜成像技术是获取大 体量神经数据的重要观测手段。然而，受限于成像时间和切片制备工艺，生物体 积数据常常面临低信噪比、大面积缺陷和高各向异性等问题，这些问题限制了 神经元分割与追踪的性能，进而影响微观脑连接图谱的重建精度。目前，神经切 片的电镜像质提升正面临三大挑战:首先，由于固定位置的电镜重复拍摄存在 畸变与漂移，难以获取成对噪声干净真值用于监督去噪。其次，生物样品制备过 程中的切片瑕疵或成像过程中的电子束击穿效应容易造成大面积内容缺陷，这 不仅阻碍了后续各向同性重建算法的应用，还降低了脑神经连接图谱的准确性。 最后，由于样品切片工艺限制，序列切片体积数据的横向物理分辨率远高于切片 厚度，这种显著的各向异性严重影响神经元分割与追踪的精度，进一步制约微观 尺度脑连接图谱的重建准确性。

为克服这些挑战，本文深入研究了电子显微镜成像的噪声模型和神经序列 切片的组织连续性等生物学先验信息。针对电镜噪声干扰、大面积缺陷以及各向 异性三大挑战，本文提出了一套系统的电镜像质提升方法，包括无损自监督去 噪、序列切片插帧以及各向同性重建。这些方法共同提升了电子显微镜图像质 量，实现了高信噪比、无缺陷和各向同性的体积数据重建，为构建高精度的微观 脑连接图谱奠定了坚实的数据基础。

  本文的主要工作和创新点归纳如下:

       1、基于掩码介质引导的隐式无损自监督去噪。针对现有电镜去噪技术所面 临的挑战，如获取成对训练数据的困难和自监督去噪方法中信号损失问题，本文 提出了一种信号无损的隐式自监督去噪框架 Blind2Unblind，解决了自监督方法 中输入信号丢失的问题，保证了电镜去噪的信号完整性。全局掩码映射器对去噪 体积中盲点处的像素进行采样，并将其映射到同一通道，实现了全局内容感知和 训练加速。为克服恒等映射问题，重新可见损失将掩码去噪辅助任务作为梯度 更新媒介，实现了盲点可见的自监督无损去噪。该框架适用于包括泊松高斯噪 声在内的空间独立噪声，在仿真和真实电镜数据集上显著优于现有自监督方法， 并表现出媲美有监督学习的去噪性能。这一成果为具有畸变漂移特性的电子显 微镜去噪提供了新的研究方向和科学去噪范式。

       2、基于细粒度噪声水平感知的定制无损去噪。本文进一步提出了一种强可 解释的显式自监督去噪框架 Blind2Sound，用于解决隐式无损去噪中生物超微结 构过度平滑和超微细节丢失的问题。自适应重新可见损失能在保持信号无损的 同时，根据噪声强度调整去噪力度以实现定制化噪声移除，避免了后处理所带来 的误差放大。通过对梯度更新中间媒介的梯度分析，本框架的训练稳定性得到了 加强。同时，克莱默高斯损失作为正则化项，促进了对噪声水平的准确感知，并 进一步提升了去噪器性能。在本框架中，作为辅助分支的噪声估计器在推理阶段被移除，有效避免了额外的计算开销。在仿真和真实电镜数据集上，该显式定制 化框架明显优于隐式无损自监督去噪，甚至略优于有监督基线，表现出强大的泛 化能力。这一研究成果进一步验证了自监督去噪范式在电镜去噪领域的突出优 势。

       3、基于全局上下文稀疏聚合的神经序列切片插帧。针对因样品制备过程中 出现的机械损伤和电子束击穿等问题所导致的电镜图像大面积缺陷，本文提出 了一种隐式表征切片间复杂形变和风格特异性的神经序列切片插帧方法。该方 法利用生物组织在 z 轴方向的连续性及全局空间上下文来建模体积数据的复杂 形变。分层稀疏自注意力聚合模块将两级稀疏分解的长短程依赖结果聚合，实现 计算量友好的隐式形变建模和全局自适应采样。为处理序列切片间的风格特异 性，本文设计的自适应风格平衡损失将 z 轴的风格上下文考虑在内，生成了真实 可靠的缺陷修复结果。实验结果证明，本方法在神经序列切片插帧任务上具有出 色的精度和鲁棒性，有望在电镜成像的大面积修复上得到大规模应用。

       4、基于多视角连续性蒸馏的体积非对称超分辨率重建。针对神经切片体积 数据的各向异性问题，本文提出了多视角生物组织连续的各向同性重建框架。首 先，多视角时空集成模块能够提取 x-z 和 y-z 平面的多尺度金字塔特征，并计算 x-y 平面的相关性系数。通过采用 x-y 平面的区域相关性采样合成预测的各向同 性中间特征，对不同轴向平面的连续性进行约束。针对误匹配中间特征带来的性 能退化，本文进一步提出轻量可堆叠的反馈蒸馏模块，用于校正中间特征，减少 大变形引起的误匹配。误匹配反馈蒸馏模块能够根据输入特征和预测的中间特 征自适应地执行反馈蒸馏，输出精确的中间插值特征。本方法在电镜非对称超分 辨率重建 CREMT 基准数据集上取得了最先进的性能，实现了生物体积数据的 近似各向同性，从而提高了神经元形态和连接关系的重建精度。

As a crucial nanoscale imaging technique, serial sectioning electron microscopy imaging is a vital observation tool for obtaining large-volume neural data. However, limitations in imaging time and slice generation technology often result in biological volume data suffering from low signal-to-noise ratios, extensive defects, and high anisotropy, which hinder the performance of neuronal segmentation and tracking, consequently impacting the reconstruction accuracy of microscopic brain connectivity mapping. Presently, there are three major challenges in enhancing the electron microscopic image quality of neural slices: First, acquiring paired noise-clean data for supervised denoising is challenging due to aberrations and drifts in repeated electron microscopic shots at fixed locations. Second, section defects during biological sample preparation or electron beam breakdown effects during imaging tend to cause large content defects, which not only impede the application of subsequent isotropic reconstruction algorithms but also reduce the accuracy of brain neural connectivity mapping. Finally, due to sample slicing process limitations, the lateral physical resolution of serial section volume data is much higher than the section thickness. This significant anisotropy severely affects the accuracy of neuronal segmentation and tracking, further limiting the reconstruction accuracy of microscale brain connectivity atlases.

To overcome these challenges, this thesis investigates biological prior knowledge, such as electron microscopy imaging noise models and tissue continuity of neural serial sections. In addressing the three major challenges of electron microscopy noise interference, large area defects, and anisotropy, this thesis presents a systematic method for electron microscopy image quality enhancement, including lossless self-supervised noise removal, serial section interpolation, and isotropic reconstruction. Collectively, these methods improve electron microscopy image quality, achieving a high signal-to-noise ratio, defect-free, and isotropic volume reconstruction, laying a solid data foundation for constructing high-precision microscopic brain connectivity atlases.

The main contributions and innovations of this thesis are summarized as follows:

1. Implicit Lossless Self-Supervised Denoising Based on Masked Medium Guidance. To address the limitations of existing electron microscopy denoising techniques, such as difficulties in obtaining paired training data and signal loss in self-supervised denoising methods, this thesis presents a lossless implicit self-supervised denoising framework, Blind2Unblind. This framework solves the input signal loss in self-supervised methods, ensuring the signal integrity of electron microscopy denoising. The global mask mapper samples pixels at blind spots in the denoised volume and maps them to the same channel, achieving global awareness and training acceleration. To overcome the identity mapping, the re-visible loss uses the mask denoising auxiliary task as a gradient update medium, enabling self-supervised lossless denoising with visible blind spots. Applicable to spatially independent noise, including Poisson-Gaussian noise, the framework significantly outperforms existing self-supervised methods on simulated and real electron microscopy datasets and exhibits performance comparable to supervised learning. The results provide a new research direction and scientific denoising paradigm for electron microscopes with aberration drift characteristics.

2. Customized Lossless Denoising Based on Fine-Grained Noise Level Perception. This thesis further proposes a highly interpretable explicit self-supervised denoising framework, Blind2Sound, to address over-smoothing and loss of ultrastructural details in biological ultrastructures caused by implicit lossless denoising. Adaptive re-visible loss enables customized noise removal by adjusting the denoising effort according to noise intensity while maintaining signal lossless, preventing error amplification caused by post-processing. The framework’s training stability is enhanced by gradient analysis of the gradient update intermediate medium. Simultaneously, the Cramer Gaussian loss, acting as a regularization term, promotes accurate noise level perception and further enhances denoiser performance. In this framework, the noise estimator as an auxiliary branch is removed during inference, effectively avoiding additional computational overhead. On both simulated and real electron microscopy datasets, this explicit customization framework significantly outperforms the implicit lossless self-supervised denoising and even slightly surpasses the supervised baseline, exhibiting strong generalization ability. This research further validates the outstanding advantages of the self-supervised denoising paradigm in electron microscopy denoising.

3. Neural Serial Section Interpolation Based on Global Context Sparse Aggregation. To address large defects in electron microscopy images caused by mechanical damage and electron beam breakdown during sample preparation, this thesis proposes a neural serial section interpolation method that implicitly characterizes complex deformation and style specificity between sections. The method leverages the continuity of biological tissues in the z-axis direction and global spatial context to model the complex deformation of volumetric data. The hierarchical sparse self-attentive aggregation module aggregates the long- and short-range dependent results of two-level sparse decomposition, achieving computationally friendly implicit deformation modeling and global adaptive sampling. To address style specificity among serial sections, the adaptive style balance loss, designed in this thesis, considers the z-axis style context and generates realistic and reliable defect repair results. Experimental results demonstrate that the proposed method offers excellent accuracy and robustness for the neural serial section interpolation task and is expected to be applied on a large scale for extensive area restoration in electron microscopy imaging.

4. Volumetric Asymmetric Super-Resolution Reconstruction Based on Multi-view Continuity Distillation. To tackle the anisotropy of volumetric data in neural slices, this thesis proposes an isotropic reconstruction framework with a multi-view biological tissue continuum. Firstly, the multi-view spatio-temporal integration module extracts multi-scale pyramidal features in x-z and y-z planes and calculates the correlation coefficients in the x-y planes. The continuity of different axial planes is constrained by synthesizing the predicted isotropic intermediate features using regional correlation sampling of x-y planes. To address the performance degradation caused by mismatched intermediate features, this thesis proposes a lightweight, stackable feed- back distillation module for correcting intermediate features and reducing mismatches caused by large deformations. The mismatched feedback distillation module adaptively performs feedback distillation based on input features and predicted intermediate features, outputting accurate intermediate interpolated features. This method achieves state-of-the-art performance on the CREMT benchmark dataset for electron microscopy asymmetric super-resolution reconstruction, attaining approximate isotropy of the bio-volume data and thus improving the reconstruction accuracy of neuronal morphology and connectivity relationships.