光学成像方法作为一种非侵入低损伤的工具，被广泛地应用到生物医学领域中。偏振光学成像对于样本亚波长微观结构敏感，能够探测到更高维度的丰富信息，在生物样本中的病理检测和定量分类中展现出强大的能力。缪勒矩阵是介质偏振性质的表征，通常使用缪勒矩阵显微镜对样本的偏振特征进行测量。在测量的过程中，需要至少四次入射光的偏振态调制来记录光与介质发生相互作用后的偏振态，并基于此计算缪勒矩阵。多次曝光会使得测量装置由于系统或者样本不稳定、图像配准失调等而引入误差与噪声，从而影响进一步使用具有明确物理意义和光学特性的偏振参数对样本进行进一步分析。本文主要研究的是基于斯托克斯成像的图像转换，针对生物医学研究中的痛点，结合人工智能技术与偏振数据的特点，提高缪勒矩阵显微镜的性能。本文研究的内容包括基于偏振斯托克斯成像的模态转换、样本虚拟染色以及多波长缪勒特征生成，实现通过缪勒矩阵显微镜的单次曝光赋予样本多种染色风格的明场对比度，并能够生成不同波长下样本的偏振参数图像，从而提高整个系统的鲁棒性。这个方法在提高采集数据效率的同时，也能够为临床医生诊断提供一种辅助手段。论文首先从成像模态的角度出发，基于深度学习的方法建立了从斯托克斯图像到明场图像的统计映射转换，实现了在不改变硬件装置和光路设计的条件下，扩大缪勒矩阵显微镜的多模态成像范围。在其基础上，又进一步实现了样本染色模态的转换，可以在单次曝光的情况下，同时输出同一样本视野之下的不同染色风格的明场对比度信息，这能够提供样本不同方面的病理情况。最后，从偏振测量本身考虑，针对在测量多波长缪勒数据时可能出现的由于对焦不准确带来误差的问题，提出了一个跨波长缪勒特征生成模型，从而实现从一种波长下采集得到的斯托克斯图像到其它多种波长下的缪勒参数图像生成，该方法能够简化硬件设计。基于偏振斯托克斯快速成像的图像模态转换有助于提高病理学家分析复杂样本不同种类数据特征的效率和稳定性。
Optical imaging methods are widely used in the biomedical field as a non-invasive and low-damage tool. Polarization optical imaging is sensitive to the sub-wavelength microstructure of the sample and is capable of detecting a wealth of information in higher dimensions, demonstrating a powerful capability in the detection and quantitative classification of pathologies in biological samples. The Mueller matrix is a representation of the polarization properties of a medium, and the polarization characteristics of a sample are typically measured using Mueller matrix microscopy. During the measurement process, at least four modulations of the polarization state of the incident light are required to record the polarization state of the light as it interacts with the medium, from which the Mueller matrix is calculated. Multiple exposures can cause the measurement device to introduce errors and noise due to system or sample instability, image misalignment, etc., which affects the further analysis of the sample using polarization parameters with well-defined physical meaning and optical properties. This paper focuses on image translation based on snapshot Stokes imaging to improve the performance of Mueller matrix microscopy by combining artificial intelligence techniques with the characteristics of polarization data to address the pain points in biomedical research.The research in this paper includes imaging modality translation, virtual staining of samples, and multi-wavelength Mueller feature generation based on polarization Stokes imaging, so as to achieve the bright-field contrast of multiple staining styles conferred on samples by a single exposure of the Mueller matrix microscope, and to be able to generate polarization parameter images of samples at different wavelengths, thus improving the robustness of the whole system. This method can provide an aid to clinicians in diagnosis while improving the efficiency of data collection.The paper firstly establishes a statistical mapping transformation from Stokes images to bright-field images based on a deep learning method from the perspective of imaging modality, which achieves the expansion of the multimodal imaging range of Mueller matrix microscope without changing the design of the hardware device and optical path. On its basis, a further transformation of the sample staining modality was achieved, allowing the simultaneous output of bright-field contrast information for different staining styles under the same sample field of view in a single exposure, which is able to provide different aspects of the sample‘s pathology. Finally, from the consideration of polarization measurement itself, to address the problem of errors due to inaccurate focusing that may occur when measuring multi-wavelength Mueller data, a cross-wavelength Mueller feature generation model is proposed so as to achieve the generation of Mueller parameter images from Stokes images acquired at one wavelength to multiple other wavelengths, which can simplify the hardware design. Image cross-modality translation based on polarization snapshot Stokes imaging helps to improve the efficiency and stability of pathologists in analyzing different kinds of data features in complex samples.
