Artificial Intelligence in Digital Holographic Imaging

17High‐throughput Label‐free Cell Counting with Deep Neural Networks

17.1 Introduction

Previous chapters introduced a digital holographic microscopy (DHM) system to reconstruct phase images of red blood cells (RBCs) using a numerical reconstruction method. This numerical reconstruction algorithm includes processes such as spatial filtering, phase unwrapping, and numerical propagations of complex diffraction waves. Since cell studies based on phase images need a numerical reconstruction step, cell analysis may benefit from this new scheme that can completely eliminate the numerical reconstruction step.

Reconstructed phase image can clearly reveal targets while those in the raw hologram are much more vague. The cell edge in the diffraction pattern recorded by DHM is usually not clearly defined. Therefore, it may be difficult to perform cell analyses based on a raw hologram using traditional image processing due to ambiguity of biological cells in the diffraction pattern. Fortunately, the diffraction pattern recorded by DHM can be analyzed using deep neural networks, which do not require defined specific features within the raw hologram. They can automatically extract useful features according to the goal of the networks.

In this chapter, we will introduce the U‐Net algorithm for cell detection and counting using the DHM diffraction pattern [1]. The U‐Net algorithm is an end‐to‐end segmentation method that uses images as input and output. It can overcome disadvantages of the CNN ...