Development of digital holographic microscopy combined with artificial intelligences

Abstract

Digital holographic microscopy (DHM) is a powerful imaging technique that encrypts the 3D information of a test sample into a single shot of 2D interference patterns. Thus, DHM has been used in various fields including examination of the biophysical properties of test samples, identification of microscale objects, and dynamic analysis of particles or cells. In this thesis, DHM technique was combined with artificial intelligences (AIs) to overcome technological limitations of the conventional DHM technique. The developed AI-based DHM technique was applied to diagnose hematological diseases and classify microparticles. Accurate and immediate identification of cell types of red blood cells (RBCs) is important for medication of several hematological diseases. Conventional methods used for classifying RBCs are time consuming and rely on the personal skill of experts. For automatic identification of RBC types, a new automatic label-free sensor was developed by combining the DHM with machine learning algorithms. Main features, including information on intensity distributions, morphological descriptors, and optical focusing characteristics, were quantitatively obtained from numerically reconstructed holographic images. All individual features for discocytes (DCs), echinocytes (ECs), spherocytes (SCs), and malaria-infected RBCs (iRBCs) are statistically different. Several machine learning algorithms were applied to largely improve the performance of cell-type identification. Among the tested algorithms, the decision tree model exhibits the best identification performance for old stored RBCs (>97%). On the other hand, the support vector machine (SVM) model shows the highest accuracy in diagnosing malaria-infected RBCs (>96%). To develop a portable and smart platform for accurate classification of particle size, the smartphone-based DHM and machine-learning algorithms were synergistically integrated. The smartphone-based DHM system consists of a coherent laser beam, a pinhole, a sample holder, a 3D printed attachment, and a modified built-in smartphone camera module. Holograms of various microparticles with different sizes were recorded with a wide field-of-view and high spatial resolution. To establish a proper classification model, tens of features including geometrical parameters and light-intensity distributions were extracted, and machine learning algorithms were applied. The SVM model trained by using three geometrical parameters and three extracted parameters from light-intensity distributions shows the highest accuracy in the classification of particles (>98%). Conventional DHM requires an expensive and bulky coherent light source and high-quality optical elements to induce diffraction and interference. To resolve these shortcomings, the deep neural network based on generative adversarial network (GAN) was utilized to perform image transformation from a defocused bright-field (BF) image acquired using a general incoherent light source to a holographic image. As a result, holograms generated from BF images through the trained GAN exhibit enhanced image contrast with 3-5 times increased signal-to-noise ratio (SNR), compared to ground truth hologram and provide valuable 3D volumetric information. Conclusively, the developed AI-based DHM technique would be potentially applied to a wide range of practical applications, such as mobile healthcare and environmental monitoring.