Artificial Intelligence in Digital Holographic Imaging

13Red Blood Cell Phase‐image Segmentation with Deep Learning

13.1 Introduction

In Chapter 12, we segmented RBC phase images obtained by DHM using the marker‐controlled watershed algorithm combined with morphological operations. However, this method cannot properly segment heavily overlapped RBCs or those touching multiple cells. Therefore, is essential to develop a more robust algorithm for RBC phase‐image segmentation.

Deep learning is a promising technique that can offer results superior to those obtained by traditional methods. Consequently, it is extensively studied in the computer vision community [1–8]. CNNs are used for image classification with great success. Recurrent neural networks provide reasonably good performance for text classification and translation. Fully convolutional neural networks (FCNs) are proposed for semantic segmentation with surprising outcomes. FCNs have the advantage of end‐to‐end training with pixel‐wise prediction. Moreover, the size of the image provided to an FCN algorithm can be arbitrary. Other FCN algorithms such as U‐net and SegNet are also suggested for semantic segmentation and are applied to biological images. In this chapter, we will apply the FCN method to RBC phase images for RBC segmentation [9]. We will introduce two RBC segmentation schemes. In the first scheme, FCN‐1, RBC phase images and manually segmented RBCs were used as a true label to train the FCN model. The trained FCN model was then applied to predict foreground (RBC) ...