br The typical automatic methods applied in image segmentati
The typical automatic methods applied in image segmentation are specific to different types of microscopy images and stain-ing technologies. The vast majority of these are based on several underlying algorithms: intensity thresholding, multifractal analy-sis, morphology operation, watershed transformation, clustering, graph-based methods, and supervised classification (Chen, Yeh, & Tzeng, 2008; Kruk et al., 2014; Reljin, Slavkovic-Ilic, Tapia, Cihoric,
& Stankovic, 2017; Soille, 2003; Xing & Yang, 2016). The methods cooperate with artificial intelligence tools, such as neural networks, support vector machines, decision trees, and Bayesian classifiers. For FISH images, the most commonly used approaches are those based on watershed transformation and morphological operations (Hsu, 2012; Raimondo et al., 2005; Theodosiou et al., 2008), as well as multifractal analysis (Reljin et al., 2017).
The common disadvantage of all of these approaches is their high sensitivity to sharp changes in the brightness intensity of the RGB channels, presence of artifacts, and inability to deal with over-lapped cells. Such deformations result in improper segmentation (for example, omission of certain cells, improper localization, and incorrect boundaries). Certain areas are often omitted in assess-ment, and the gain amplification ratio is affected as a result. All of these problems create the need to develop methods for proper cell reconstruction based on their distorted shapes. Thus far, this problem has not been addressed satisfactorily. In this study, we propose a reconstruction procedure that allows for discovering de-formed cells, identifying the area of artifacts, correcting the shape of the deformed cells, and finally, obtaining a proper set of KRN 7000 in the analyzed field of view.
The main novel contributions of this work include an auto-matic method for reconstructing the cell nuclei with an acceptable accuracy level. The randomized correspondence algorithm, known as PatchMatch (Barnes, Shechtman, Finkelstein, & Goldman, 2011) that was applied in the solution was enriched by two main factors: specially defined sensitivity and similarity measures. Their applica-tion in the reconstruction procedure allows for accelerated compu-tations and an increased probability of obtaining proper segmenta-tion results.
Although the PatchMatch approach has been used in video technology for object detection, label transfer, or detecting digital forgeries (Barnes et al., 2011), it has never been applied to cell seg-mentation. Its only biomedical application has been found in car-diac image processing (Shi et al., 2014), which is a significantly dif-ferent problem from cell segmentation. The implementation of the presented procedure can lead to a significant improvement in the quality of cell segmentation, and as a result, increase the accuracy of the HER2/CEN17 ratio in FISH images.
The proposed technique allows for automatic reconstruction of the preliminary segmented cells and their recovery in a corrected form, which is acceptable in medical practice. The reconstructed cell nuclei outline is very close to its real boundary. Moreover, it enables the level of cell overlapping to be assessed, which is a fac-tor that significantly influences the HER2/CEN17 ratio in medical examinations. The results of the numerical experiments demon-strate that the reconstruction quality of cells is comparable to the results obtained by experts, while the repeatability, precision, and
duration of the operation are definitely more beneficial for the pro-posed computer system.
The remainder of this paper is organized as following. The problem statement is provided in Section 2. A basic detailed de-scription of the method is presented in Section 3. In Section 4, the results of the numerical experiments are presented and discussed. Conclusions and future research are outlined in the final Section 5.
2. Problem statement
FISH is a cytogenetic technique that allows for detecting and vi-sualizing specific DNA sequences in the tissue material. Fluorescent dyes are applied to the tissues, which emit their own light with a specific color under the influence of an external light. A marked gene (biomarker) may be visible as a red spot (HER2) or green spot (CEN17). Moreover, cell nuclei exist that are visible as blue or purple areas. Pathologists analyze the number of HER2 genes and chromosome centromeres identified as CEN17 in the separated cells, and define the HER2 gene amplification classes according to their HER2/CEN17 ratio, as follows: positive when the ratio is higher than 2.2, negative when the ratio is smaller than 1.8, and dubious when its value is in the range of 1.8–2.2 (Theodosiou et al., 2018; ER2FISH, 2014; Lopez et al., 2013; Masmoudi et al., 2009).
Unfortunately, in practice, large parts of cells are overlapped by stroma or other biological material, visualized as a green-yellow color. The remainder of the image is filled with a dark, gray-black intercellular space. A typical example of the entire field of view, visible under a microscope with 100× lens magnification and 10× internal magnification, is presented in Fig. 1.