Accurate segmentation of touching cells in multi-channel microscopy images with geodesic distance based clustering

Abstract

Multi-channel microscopy images have been widely used for drug and target discovery in biomedical studies by investigating morphological changes of individual cells. However, it is still challenging to segment densely touching individual cells in such images accurately and automatically. Herein, we propose a geodesic distance based clustering approach to efficiently segmenting densely touching cells in multi-channel microscopy images. Specifically, an adaptive learning scheme is introduced to iteratively adjust the clustering centers which can significantly improve the segmentation accuracy of cell boundaries. Moreover, a novel seed selection procedure based on nuclei segmentation is suggested to determine the true number of cells in an image. To validate this proposed method, we applied it to segment the touching Madin-Darby Canine Kidney (MDCK) epithelial cells in multi-channel images for measuring the distinct N-Ras protein expression patterns inside individual cells. The experimental results demonstrated its advantages on accurately segmenting massive touching cells, as well as the robustness to the low signal-to-noise ratio and varying intensity contrasts in multi-channel microscopy images. Moreover, the quantitative comparison showed its superiority over the typical existing cell segmentation methods.

Introduction

High content screening (HCS) has been widely used for discovering novel drugs and targets by investigating the morphological changes of interested proteins inside individual cells [1], [2], [3]. Along with the advance of automated image acquisition equipments, large scale of cell images with multiple fluorescent markers is being generated in experiments. Whereas, great challenges have been posed on the automated quantification of individual cell morphology due to their complex appearances, uneven intensity, low signal-to-noise ratio (SNR), and cell touching [4], [5], [6], [7]. Automated detection and accurate segmentation of touching cells are crucial for biological studies such as cell morphological analysis, cell tracking and cell phase identification [8], [9], [10]. However, it remains being an open problem due to the aforementioned challenges.

A series of approaches of automated cell segmentation in different kinds of images have been reported [11], [12]. In general, the widely used segmentation approaches, for example, watershed method [9], [13] and deformable models like the snake model and level set methods [14], [15], are sensitive to initializations (e.g., seeds selection), and intensity variation inside cells. Specifically, the watershed method is often involved in the over-segmentation problem, and also sensitive to the intensity noise. Active contour or level set based deformable models are very sensitive to boundary initializations. In addition, the computational cost is heavy for these methods. Therefore, they are unsuitable for processing a large scale image dataset in high content screening studies. Clustering methods and statistical approaches have also been employed for cell segmentation [16], [17], [18]. These methods are efficient but usually provide incomplete segmentations because they operate on the single pixel level and neglect the fact that cells are continuous regions. Especially, these methods have limited ability in delineating boundaries of touching cells in cell membrane segmentation. Moreover, large intensity variation on cell membrane and inside cells has made it a much more challenging problem to find the dividing boundaries of cells, which usually lead to biased segmentation results.

In order to segment touching cells more accurately, several new techniques have been investigated in recent years, including certain adaptive learning approaches to separating overlapped nuclei or cells. In fact, Jung et al. [19] utilized an unsupervised Bayesian classification scheme based on parametric EM algorithm for separating overlapped nuclei. Moreover, the single-path voting followed by a mean-shift clustering process was employed to locate the cell centers by Qi et al. [20]. Furthermore, the graph-cuts-based algorithm was successfully applied by Yousef et al. [21] to develop a nuclei segmentation method as well as a powerful toolkit software named ”FARSIGHT”. On the other hand, some classical conventional methods have also been carried forward, such as the adaptive thresholding watershed algorithm [9] and the adaptive active physical deformable model [22]. To solve the cell segmentation problem more effectively via the advanced mathematical model, Jones et al. [23] proposed a new distance metric defined on image manifolds which combined image gradients and inter-pixel distance together. This metric demonstrated its effectiveness, but the segmentation results strongly depended on the initial selection of seeds. Although the centers of nuclei could be adopted as the seeds in CellProfiler [24], the segmentation results became worse as cell nuclei are touched together. Besides, CellProfiler accomplished the pixel clustering after only one time of assignment, the boundaries were obtained without optimization. Along the direction of new distance metric, we recently proposed a novel clustering method with the geodesic distance for gray scale images in [25] with little a priori knowledge. Although it is robust to the initialization of seeds, the determination of number of clusters is still a bottleneck in improving the segmentation performance further. Over-segmentation occurs when the number of initial seeds is more than the true number of cells in an image, which is a well known inherent disadvantage of K-means clustering methods.

Aiming for accurate segmentation of the touching Madin-Darby Canine Kidney (MDCK) epithelial cells in multi-channel (color) images, in this paper, we propose a nuclei segmentation based K-means clustering method using the geodesic distance. The main contribution of this paper is that an adaptive learning scheme is established to adjust the clustering centers with the geodesic distance for improving the segmentation accuracy. Technically, color information is taken into consideration to select the number of reasonable initial seeds for clusters through a well-designed nuclei segmentation procedure. As a result, the over-segmentation problem is effectively overcome, more accurate dividing boundaries are delineated, and the convergence of the improved K-means clustering method is also speeded up.

The rest of this paper is organized as follows. We briefly introduce the geodesic distance based clustering method in Section 2. Section 3 presents the details of our adaptive cell segmentation procedure. The experimental results on MDCK images are demonstrated in Section 4. Finally, a brief conclusion is provided in Section 5.