A visual targeting system for the microinjection of unstained adherent cells

Abstract

Automatic localization and targeting are critical steps in automating the process of microinjecting adherent cells. This process is currently performed manually by highly trained operators and is characterized as a laborious task with low success rate. Therefore, automation is desired to increase the efficiency and consistency of the operations. This research offers a contribution to this procedure through the development of a vision system for a robotic microinjection setup. Its goals are to automatically locate adherent cells in a culture dish and target them for a microinjection. Here the major concern was the achievement of an error-free targeting system to guarantee high consistency in microinjection experiments. To accomplish this, a novel visual targeting algorithm integrating different image processing techniques was proposed. This framework employed defocusing microscopy to highlight cell features and improve cell segmentation and targeting reliability. Three main image processing techniques, operating at three different focus levels in a bright field (BF) microscope, were used: an anisotropic contour completion (ACC) method, a local intensity variation background-foreground classifier, and a grayscale threshold-based segmentation. The proposed framework combined information gathered by each of these methods using a validation map and this was shown to provide reliable cell targeting results. Experiments conducted with sets of real images from two different cell lines (CHO-K1 and HEK), which contained a total of more than 650 cells, yielded flawless targeting results along with a cell detection ratio greater than 50%.

Introduction

Progress in biotechnology has been increasing the demand for genetically modified organisms [1], with manual microinjection forming one of the most widely used techniques to deliver foreign DNA or other proteins into cells. These operations are typically characterized by high skill requirements, long training periods (1–2 years) and comparatively low success rates (40%–70%). They are also considered tedious and time consuming [2], [3], [4]. Moreover, humans lack consistency over repeated operations, so the automation of cell microinjections is a growing field of interest with a wide range of applications [1], [2], [3], [4]. In this context automatic cell segmentation is essential to enable the realization of automated systems that can increase the efficiency and consistency of the operations while reducing costs.

Historically, different techniques have been used to ease the visual identification of cells under the microscope. Most of these rely on the use of chemical dyes [5], [6], [7]. Regardless of the results achieved with staining techniques, side effects of chemical dyes on the cells' life expectancy are still under debate. Several works report on the toxicity of the dyes describing side effects observed, for example, in the endoplasmatic reticulum [8] and cytoskeleton [9]. Because of this potential harm highly diluted dye solutions have been used for a long time, impairing the quality of the staining process [10]. Moreover, staining may affect cell membrane permeability or dye uptake may differ between cell types in the case of multiline cell cultures [11].

In contrast, the new vision algorithm presented here is designed to detect unstained cells using bright field (BF) microscopy. This chemical dye free technique avoids the controversial side effects mentioned above and is based on a widely used imaging mode for cells observation and microinjection.

The research presented here is mainly focused on Chinese hamster ovarian cells (CHO-K1) imaged using 200× optical magnification and 1024×768 pixels resolution. These cells are widely used in biological and medical research, and also commercially in the production of therapeutic proteins [12]. They are mostly transparent, adherent to the Petri dish and present a range of different shapes and sizes with dimensions varying from 10 to 50 μm (see Fig. 1). In addition, human embryonic kidney (HEK) cells were also used in this research as part of the experimental validation of the reported visual targeting framework.

Literature in the field of unstained cell targeting shows that pioneering image processing algorithms for cell segmentation were often based on gray level thresholding [13], however, these methods fail to single out individual cells in clusters. Nevertheless, most state of the art methods are still threshold-based, especially when the aim is a coarse preliminary selection of the targets [14], [15]. Recently, Tscherepanow et al. [15] applied constrained active contours (snakes) for surface detection of plated ovarian cells achieving a recognition rate up to 90%. However, snakes are prone to the effects of local minima during the energy function minimization process, which may trap the snakes in wrong models [16].

A completely different approach focuses on the cell membrane in both 2D and multilayered 3D representation. Some cells show very distinctive outer contours that can be easily detected by edge detection and ridge enhancement algorithms. Examples that exploit this characteristic include the work of Adiga et al. [18], in which partial derivative equations (PDE) based edge enhancing diffusion [19], [20] was used to achieve better smoothing along the edges than across regions, resulting in up to 95% correct nuclei segmentation and false positive rates of around 10%. In this case the diffusion process assisted an anisotropic contour completion (ACC) process [17] that has demonstrated good results when applied to CHO-K1 cells [21], [22].

Another useful approach to cell detection is based on supervised learning schemes. Given a training set of 496 cells, Long et al. [23] used a support vector machine to count viable cells in BF microscope images and compared this count with the results obtained from a neural network [24], [25]. Both methods resulted in a very high success rate (up to 97%). This is one of the best performances yet reported, but the method is limited to cells that are visible and show features that can be compared to the template image worked out in the training set.

Ali et al. [16] used defocused BF microscope images to obtain a phase-based segmentation of unstained cells. They started with a highly defocused image in which the cell was visible as a strong dark smear and proceeded by searching its outer contours as focus was improved. As the presence of visual high frequency elements increased with the improving focus, they used local phase coherence to refine the cell contour, thus refining the localization of its boundaries. The result was a cell detection rate consistently above 87% with false positive rates down to 5%. Unfortunately, despite these good results, this method was only developed to segment isolated cells and is not directly applicable to the cells of interest here since CHO and HEK cells grow in tight clusters.

To achieve reliable CHO and HEK cells targeting, this work proposes the use of defocusing as a means to improve cell contrast and cell localization performance. Different focus levels highlight different cell features, hence the developed framework implements distinct image processing algorithms to best exploit the information provided at different focus levels (Fig. 2). Increased consistency and reliability is achieved by combining the output of these algorithms through validation maps. Supporting this theory is the fact that defocusing BF microscopy has been recently considered in several works on automatic cell spotting due to it being simple, affordable and suitable for an objective comparison between different frameworks. Most of these works rely on the use of several focus levels as in Selinummi et al. where the authors collected a z-stack of each frame displayed to generate a 2D projection image [26] or in Ali et al. where the information coming from in-focus and out-of-focus images is combined to spot adherent cells' nuclei [27]. Watershed-based cell spotting techniques have also been tried over defocused images with some success [28]. In addition, when the microscope focus cannot be directly set, other methods employ progressive coarsening over the in-focus images to simulate a blurring effect [29].

Differently from most of the literature, the main goal of our framework was the achievement of an error-free targeting system, which is important to guarantee high consistency in automatic microinjection procedures. This means defining only one target per cell. It also means that maximizing cell identification rate was a secondary goal defined to target a large percentage of the cells in the field of view while minimizing waste of injection material.

In the following sections this new processing framework is thoroughly described, starting with a review of defocusing theory and of the main image processing algorithms used at each focus level. The fusion of information for enhanced cell targeting results is subsequently presented, followed by evaluation and validation experiments conducted with the two adherent cell lines mentioned above. Lastly, conclusions and future applications are presented.