[论文解读] Artificial confocal microscopy for deep label-free imaging

Xi Chen, Mikhail E. Kandel+9 more|arXiv (Cornell University)|Oct 28, 2021

Digital Holography and Microscopy参考文献 46被引用 6

一句话总结

本文提出人工共聚焦显微成像（ACM），一种深度学习方法，通过训练卷积神经网络（CNN）从定量相位成像数据预测共聚焦荧光样图像，实现高对比度、深度分辨、无标记成像。该方法在未标记的生物样本（如微球、神经元和3D类器官）中实现了共聚焦级的切片效果，可实现无创的细胞核分割与干质量定量，且无光毒性。

ABSTRACT

Widefield microscopy methods applied to optically thick specimens are faced with reduced contrast due to spatial crosstalk, in which the signal at each point is the result of a superposition from neighboring points that are simultaneously illuminated. In 1955, Marvin Minsky proposed confocal microscopy as a solution to this problem. Today, laser scanning confocal fluorescence microscopy is broadly used due to its high depth resolution and sensitivity, which come at the price of photobleaching, chemical, and photo-toxicity. Here, we present artificial confocal microscopy (ACM) to achieve confocal-level depth sectioning, sensitivity, and chemical specificity, on unlabeled specimens, nondestructively. We augmented a laser scanning confocal instrument with a quantitative phase imaging module, which provides optical pathlength maps of the specimen on the same field of view as the fluorescence channel. Using pairs of phase and fluorescence images, we trained a convolution neural network to translate the former into the latter. The training to infer a new tag is very practical as the input and ground truth data are intrinsically registered and the data acquisition is automated. Remarkably, the ACM images present significantly stronger depth sectioning than the input images, enabling us to recover confocal-like tomographic volumes of microspheres, hippocampal neurons in culture, and 3D liver cancer spheroids. By training on nucleus-specific tags, ACM allows for segmenting individual nuclei within dense spheroids for both cell counting and volume measurements. Furthermore, taking the estimated fluorescence volumes, as annotation for the phase data, we extracted dry mass information for individual nuclei. Finally, our results indicate that the network learning can be transferred between spheroids suspended in different media.

研究动机与目标

• 为克服宽场显微镜在光学厚样品中因空间串扰导致图像对比度下降的局限性。
• 消除深层组织成像中对荧光标记和光毒性的依赖。
• 开发一种非侵入性、高对比度的成像方法，实现未标记样本的共聚焦级深度切片。
• 在无需基因或化学标记的情况下，实现在3D类器官中对干质量的定量测量与细胞核分割。
• 展示训练模型在不同类器官培养基之间的可迁移性。

提出的方法

• 通过在激光扫描共聚焦显微镜上增加定量相位成像（QPI）模块，同步获取同一视场内的相位与荧光图像。
• 利用相位与荧光图像对训练卷积神经网络（CNN），仅从相位数据预测荧光样图像。
• 采用内在配准的输入（相位）与真实标签（荧光）数据端到端训练网络，实现自动且精确的图像转换。
• 训练后的模型可生成具有增强深度切片能力的人工共聚焦样图像，即使在厚而散射的样本中亦可实现。
• 利用预测的荧光体积作为伪标注，从密集的类器官中提取干质量并分割单个细胞核。
• 通过在不同培养基中培养的类器官上测试同一网络，验证了模型的可迁移性，表明其具有稳健的泛化能力。

实验结果

研究问题

• RQ1 深度学习模型能否从定量相位图像预测共聚焦样荧光图像，从而在厚而未标记的样本中实现增强的深度切片？
• RQ2 预测的荧光体积能否在无标记的情况下实现对密集3D类器官中单个细胞核的准确分割？
• RQ3 能否仅从相位数据推导出的估计荧光体积中实现干质量的定量测量？
• RQ4 训练后的模型是否能在不同生物环境（如不同类器官培养基）中实现泛化？
• RQ5 该方法能否在无光漂白或化学标记的情况下实现共聚焦级的对比度与分辨率？

主要发现

• ACM生成的图像在深度切片能力上显著优于传统宽场或相位成像，可清晰呈现微球和3D类器官内部结构。
• 该方法成功分割了密集3D肝癌类器官中的单个细胞核，实现了精确的细胞计数与体积测量。
• 利用从相位数据推导出的估计荧光体积，可高精度地量化单个细胞核的干质量。
• 训练后的模型在不同类器官培养基之间表现出良好的可迁移性，表明其对不同生物条件具有稳健的泛化能力。
• 该方法在未标记样本中实现了类共聚焦的断层成像，且无光毒性、化学标记或光漂白现象。