Cell Segmentation in Quantitative Phase Images with Improved Iterative Thresholding Method

Quantitative Phase Imaging (QPI) is a label-free microscopic technique, which provides images with high contrast, moreover, it provides quantitative cell mass measurements for each pixel. Segmentation of particular cells is an important step in the analys

  • PDF / 1,635,036 Bytes
  • 7 Pages / 439.37 x 666.142 pts Page_size
  • 28 Downloads / 197 Views

DOWNLOAD

REPORT


Abstract. Quantitative Phase Imaging (QPI) is a label-free microscopic technique, which provides images with high contrast, moreover, it provides quantitative cell mass measurements for each pixel. Segmentation of particular cells is an important step in the analysis of QPI image data. This paper presents a method for automatic cell segmentation in QPI images. The proposed method improves iterative thresholding, which is a very promising method, however, it is not able to segment densely clustered cells. Our improved iterative thresholding includes two additional steps – Laplacian of Gaussian image enhancement and distance transform-based splitting. The method was compared with original iterative thresholding and another method on two cell lines, where the proposed method successfully deals with a densely clustered type of cells and achieves significantly better results on both datasets. Keywords: Cell segmentation · Quantitative Phase Imaging Iterative thresholding · Laplacian of gaussian

1

·

Introduction

Image segmentation is a very important step of microscopic cell image analysis, which enables the extraction of parameters of individual cells and, despite its extensive research, it is still an unresolved problem. This work focuses on instance segmentation, where individual cells are separated, compared to semantic segmentation, where the goal is to classify pixels as cells/background [13]. A large variety of cell segmentation methods exist, where many methods are based on watershed transform [5] or include Laplacian of Gaussian (LoG) filtering [2] for cell detection as one of the steps (see reviews [13,14] for more details). Currently, best performing methods are deep-learning-based, specifically most of them is based on U-Net segmentation architecture [6] with specifically designed network output in order to be able separate individual cells (e.g. StarDist [7] and CellPose [10]). Despite its superior results, deep-learning methods require a c Springer Nature Switzerland AG 2021  T. Jarm et al. (Eds.): EMBEC 2020, IFMBE Proceedings 80, pp. 233–239, 2021. https://doi.org/10.1007/978-3-030-64610-3_27

234

T. Vicar et al.

large amount of training data, thus standard methods are still required for fast segmentation of small datasets. Quantitative Phase Imaging (QPI) is a label-free microscopic technique, which provides images with high contrast, moreover, it provides quantitative mass measurements for each pixel, which enables to measure cell dry mass and its distribution in cells. Nevertheless, there are only a few specialized methods for the segmentation of QPI images, which uses its specific properties. Loewe et al. [3] described a simple and efficient method for QPI cell segmentation, which is based on iterative thresholding with increasing threshold, where individual objects created with larger threshold are used only if they have a mass larger then selected mass limit. However, similarly to watershed-based techniques (e.g. Q-PHASE’ dry mass guided watershed in [13]), it is not able to separate individual cel