Adaptive filtering and hypothesis testing: Application to cancerous cells detection

Abstract

We propose a new method to detect cells in microscopic imagery, the problem under study being the analysis of cancerous cells experiencing metastasis i.e. cells susceptible to migration. This work would help medical researchers to study the evolution of a cancer. The peculiar nature of the images due to the acquisition protocol causes some difficulties. These are resolved through tailored preprocessing involving correction of uneven illumination and enhancement of cellular information. Detection and counting of cells are performed by our proposed filtering that provokes peaks in its convolution space wherever cells are present. We compare our counting results with those provided by human experts and with a Hough transform developed for similar purposes. The validity of the cell segmentation from the peaks is then established by a statistical test of the closeness of the segmented cell to a cell model.

Introduction

This work was motivated by a question concerning migration of cancerous cells (Vigneron et al., 2007). These cells are clones of the cell that initiated the cancer, having acquired certain characteristics allowing it to divide indefinitely and be able to metastase, i.e. to proliferate and migrate. Cell migration is of two types: mesenchymatic and amoeboid (Friedl and Wol, 2003). Amoeboid migration is fast and is usually responsible for metastasis, while mesenchymatic migration results in proliferation within the same tumour. During their migration process, amoeboid cells first adopt an amoeboid form referred to as blebbing. This allows them fast movement governed by weak and short-term interactions with the extracellular matrix. Due to their high deformability, they can slip through the extracellular matrix without deteriorating it. On the contrary, mesenchymatic cells become polarized and choose a migration front along which form new points of adhesion to the extracellular matrix, while those on the posterior part of the cell detach themselves from it, propelling the cell along.

Changes in migratory behaviour have been observed experimentally and associated with phenotypic or morphological switching in various situations, such as the migration-proliferation dichotomy of the epithelial-mesenchymal transition or the mesenchymal–amoeboid transition of cancer cells in the extracellular matrix (Gille et al., 2005, Chauviere et al., 2009). Fig. 1 shows the three morphologies: blebbing (Fackler and Grosse, 2008) and splayed, pertaining to amoeboid and mesenchymatic migrations respectively as well as smooth round pertaining to the intermediate stage. The understanding of this migration process could help to arrest the development of cancer and increase the chances of a cure.

In an experiment, the goal is to determine which cells have begun the metastatic process and which have not. To do this requires counting the proportion of each type of cells and comparing these proportions. This can be done manually by an expert, but is hard and tedious. Our goal is to develop automatic image processing techniques that could replace the human intervention, while attaining comparable accuracy. This is a particularly useful tool for biologists studying the evolution of cancer under varying environmental conditions.

In this paper we will focus on cell detection. By cell detection we refer to a process that provides us the number and the locations of cells. The main focus of our work, directed toward achieving the best possible cell detection performance, is to develop a filtering step allowing optimal adaptive estimation from the local clutter. This process of cell detection and validation has consequences on the statistics obtained on the three classes of cells, as we would like to let an equal proportion of cells in each class to escape detection. Cell detection is the first step of a larger chain of processes, including cell segmentation and cell classification, the latter all depending on the accuracy of this primordial step.

Work on the detection and counting of cells in microscopy images is very varied but has mostly been focused on segmentation of cells leading as a byproduct to an automatic count. To begin with, hardware methods (Sergent-Tanguy et al., 2003, Edwards et al., 2009) exist to identify and quantify sections of cells cultured in suspension. However, being integrated into the material, they are monetarily expensive and require a trained technical specialist.

Several researchers have been developing automated methods for segmenting and counting cells in microscopy images (Sheikh et al., 1996, Zimmer et al., 2002, Garrido and de la Blanca, 2000, Anoraganingrum, 1999). Anoraganingrum (1999) performed edge detection on melanoma cells using median filter and mathematical morphology. Garrido and de la Blanca (2000) approximated red blood cell locations using a parametric ellipse model and refined its contours using a deformable template. Sheikh et al. (1996) proposed a method of identifying the major blood cell types using median and edge enhancing filters and to classify them based upon their morphological features using neural networks. Zimmer et al. (2002) suggested tracking of motile cells using a parametric active contour model, along with a comprehensive strategy of working with cellular images. Some approaches are based on machine learning (Long et al., 2005, Guggenbergera et al., 2009, Markiewicz et al., 2006, Zheng et al., 2004). Long et al., 2005, Zheng et al., 2004 proposed methods based on neural network. Markiewicz et al. (2006) proposed a method to cell recognition and count using Support Vector Machine. In this kind of approach, the major task is to create the learning set, which is usually done manually by an independent expert for cell type and is time consuming. Another disadvantage is the time spent on training and parameter adjustments. Approaches that use classical segmentation methods, such as threshold, morphological filtering and watershed transformation (Ambriz-Colin et al., 2006, Niemisto et al., 2004, Glory et al., 2004, Thorlin et al., 2009) also have been proposed. Korzynska (2008) presented a method for automatic counting of neural stem cells growing in cultures which is performed in two steps: a segmentation step where the image is separated in several regions and a counting step where each extracted region is counted as a single cell. Here, as is the case usually, counting and detection are obvious byproducts of image segmentation.

Blood smear image analysis has been tackled by using conventional image processing techniques like morphology (DiRuberto et al., 2002), edge detection (Sio et al., 2007), region growing (Theerapattanakul et al., 2004), etc., which all have shown certain degrees of success with respect to the used data.

Sio et al. (2007) addressed the problem of parasitemia estimation using edge detection and splitting of large clumps made up from erythrocytes. The outcome of the approach was shown to be satisfactory for well-stained samples with well-separated erythrocytes. For the same problem, watershed transform (Vincent and Soille, 1991) were also employed, given that local maxima indicate the centers of convex shapes, i.e. blood components particularly erythrocytes. This concept, however, is only justifiable for images which exhibit a small degree of cell overlap.

The particular nature of our images, as described in Section 2 dictates our choice of method. The data does not lend itself to correct segmentation into connected components containing potentially one cell each. Therefore, direct cell segmentation is likely to produce incoherent results and has to be preceded and therefore guided by this in-place cell detection. A similar approach has also been taken by Pinzón et al. (2004) suggested that the problem of erythrocyte segmentation could be reduced to peak selection in the Hough circle space. The study focused on detecting erythrocytes of circular shape and uniform size.

Matched filtering (Turin, 1960) could be used to construct such a search space since it maximizes the signal-to-noise ratio between a template and parts of the target image, thereby producing a response space similar to Hough with peaks corresponding to regions of the image that are most similar to the template. Locating these peaks amounts to the aforementioned cell detection.

As the remaining discussion shows, we elaborate cell detection into three steps: cell detection proper, extraction of the detected cell segment, and finally the validation of this segment as indeed being a cell. The following is organized as follows: Section 2 presents the microscopy data and discusses the particular challenges that it poses as well as the preprocessing stage developed to mitigate them. Section 3 begins by describing our cell detection approach ans discusses its advantages and its shortcomings. The following two sections present how we redress these shortcomings. Section 4 presents the segmentation algorithm used, and Section 5 presents our statistical testing method to validate the conformity of the extracted objects to cells. Finally, Section 6 concludes the paper.

A simplified scheme in Fig. 2 gives an overview of the work.