Jiao Yuheng, He Yuchen R, Kandel Mikhail E, Liu Xiaojun, Lu Wenlong, Popescu Gabriel
Quantitative Light Imaging Laboratory, Department of Electrical and Computer Engineering, Beckman Institute for Advanced Science and Technology, University of Illinois at Urbana-Champaign, Urbana, Illinois 61801, USA.
State Key Laboratory of Digital Manufacturing Equipment and Technology, School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan 430074, China.
APL Photonics. 2021 Apr;6(4). doi: 10.1063/5.0041901. Epub 2021 Apr 6.
Quantitative phase imaging (QPI) has been widely applied in characterizing cells and tissues. Spatial light interference microscopy (SLIM) is a highly sensitive QPI method due to its partially coherent illumination and common path interferometry geometry. However, SLIM's acquisition rate is limited because of the four-frame phase-shifting scheme. On the other hand, off-axis methods such as diffraction phase microscopy (DPM) allow for single-shot QPI. However, the laser-based DPM system is plagued by spatial noise due to speckles and multiple reflections. In a parallel development, deep learning was proven valuable in the field of bioimaging, especially due to its ability to translate one form of contrast into another. Here, we propose using deep learning to produce synthetic, SLIM-quality, and high-sensitivity phase maps from DPM using single-shot images as the input. We used an inverted microscope with its two ports connected to the DPM and SLIM modules such that we have access to the two types of images on the same field of view. We constructed a deep learning model based on U-net and trained on over 1000 pairs of DPM and SLIM images. The model learned to remove the speckles in laser DPM and overcame the background phase noise in both the test set and new data. The average peak signal-to-noise ratio, Pearson correlation coefficient, and structural similarity index measure were 29.97, 0.79, and 0.82 for the test dataset. Furthermore, we implemented the neural network inference into the live acquisition software, which now allows a DPM user to observe in real-time an extremely low-noise phase image. We demonstrated this principle of computational interference microscopy imaging using blood smears, as they contain both erythrocytes and leukocytes, under static and dynamic conditions.
定量相位成像（QPI）已广泛应用于细胞和组织表征。空间光干涉显微镜（SLIM）是一种高度灵敏的QPI方法，因其部分相干照明和共光路干涉测量几何结构。然而，由于四帧相移方案，SLIM的采集速率受限。另一方面，诸如衍射相显微镜（DPM）等离轴方法允许单次QPI。然而，基于激光的DPM系统受散斑和多次反射导致的空间噪声困扰。在并行发展中，深度学习在生物成像领域被证明具有价值，特别是因其能够将一种对比度形式转换为另一种。在此，我们提出使用深度学习，以单次拍摄图像为输入，从DPM生成合成的、SLIM质量且高灵敏度的相位图。我们使用一台倒置显微镜，其两个端口连接到DPM和SLIM模块，以便我们能够在同一视野获取两种类型的图像。我们构建了一个基于U-net的深度学习模型，并在1000多对DPM和SLIM图像上进行训练。该模型学会了去除激光DPM中的散斑，并在测试集和新数据中克服了背景相位噪声。测试数据集的平均峰值信噪比、皮尔逊相关系数和结构相似性指数分别为29.97、0.79和0.82。此外，我们将神经网络推理应用于实时采集软件，现在DPM用户可以实时观察到极低噪声的相位图像。我们使用血涂片在静态和动态条件下展示了这种计算干涉显微镜成像原理，因为血涂片中同时包含红细胞和白细胞。
