Automated segmentation of tissue images for computerized IHC analysis

Abstract

This paper presents two automated methods for the segmentation of immunohistochemical tissue images that overcome the limitations of the manual approach as well as of the existing computerized techniques. The first independent method, based on unsupervised color clustering, recognizes automatically the target cancerous areas in the specimen and disregards the stroma; the second method, based on colors separation and morphological processing, exploits automated segmentation of the nuclear membranes of the cancerous cells. Extensive experimental results on real tissue images demonstrate the accuracy of our techniques compared to manual segmentations; additional experiments show that our techniques are more effective in immunohistochemical images than popular approaches based on supervised learning or active contours. The proposed procedure can be exploited for any applications that require tissues and cells exploration and to perform reliable and standardized measures of the activity of specific proteins involved in multi-factorial genetic pathologies.

Introduction

In this paper we address the problem of immunohistochemical tissue image segmentation.

In the last few years biological imaging has definitely undergone a revolution and new techniques have been shown to be effective in extracting clinical and functional information from images of molecules and tissues. In particular, the quantification of the activity of specific proteins through the analysis of tissue images provides critical information about important multi-factorial genetic pathologies, such as cancer. For example, EGFR/erb-B family of receptors plays an important role for non-small cell lung carcinoma (NSCLC) development. Quantifying and classifying the EGFR expression and activity in NSCLC with special regard to the assessment of the prevalence of EGFR mutations as well as to ligand–receptor interactions lead to new insights into the modulation of EGFR in individual lung carcinomas [1].

In addition to its well-established role in the early diagnosis of cancerous diseases, protein expression analysis recently acquired a new important role in the design of novel targeted therapies: the correlation of standardized measurements of protein expression from pathological tissue images with genetic expression data extracted from the same tissues allows to define a group of potential candidates to protein family-inhibiting therapies as well as to classify the specific pathology in a more accurate way through its specific genetic alterations [1], [2], [3].

A widespread technique in this field which leads to protein activity quantification is immunohistochemistry (IHC), that is the analysis of cancer tissue images where marked antibodies are used to link specific proteins in situ, as well as their ligands; the evaluation of the colored stains at the specific subcellular regions where the markers are localized (i.e. nucleus, cellular membrane, cytoplasm) provides important information for the assessment of cancer [1]. Fig. 1 shows three images taken from different tissue specimens, where brown and blue stained regions indicate that the relative marker is present.

In the last few years this technique has acquired a central role in pathology thanks to its several advantages over alternative bioimaging techniques (e.g. fluorescence in situ hybridization, FISH); among them, its wide availability, low cost, easy and long preservation of the stained slides [4].

Currently, pathologists usually resort to slow, manual analysis to extract information from IHC images. This is time-consuming, not scalable for very large datasets and extremely affected by inter- as well as intra-operator variability [5]. Moreover, IHC’s new role in target therapy response prediction and in the correlation of protein and genetic expression data [1], [2], [3] is placing new demands on the reproducibility of the obtained information [6]. This calls for new automated tools able to assist the pathologist in their examinations and to provide fast, accurate and standardized measures of protein activity in IHC images.

Protein expression has to be computed within the specific subcellular location of interest of the analyzed receptors (nuclei, cellular membranes or cytoplasm, depending on the receptor) in order to obtain sensible biological information [1]; in fact the subcellular location in which the receptors are expressed may represent a very small percentage of the whole imaged area. This translates into a cell-by-cell evaluation of the subcellular stain, that is much more specific than simple quantification of the global amount of the stain in the whole image [7], [8] and that requires preventive recognition of the cancer cells in the specimen and accurate segmentation of the subcellular compartments where the target receptors are localized.

In this paper we present fully automated techniques that exploit: (i) recognition of the cancerous cells and separation from the non-cancerous tissue areas in the specimen (e.g. stroma) and (ii) segmentation of cancerous nuclei in the areas identified at point (i). The proposed procedure can be exploited to perform reliable and standardized measures of protein activity as well as for any other applications that require tissue and cell exploration.

To this day, literature does not provide effective fully automated procedures to perform the tasks addressed by this paper. Commercial microscopy software usually allow the user to select the subcellular location of interest for protein activity quantification in a semiautomated or automated way; nevertheless, these tools are based on simple approaches that suffer due overgeneralization, and most of the times require extensive user interaction, so that the objectivity of the result tends to be lost [4], [6], [9], [10]. In particular, commercial products generally require the user to select manually the areas that are richest in tumoral cells [11], [12], to set intensity thresholds or levels to distinguish the cellular patterns from the background [13], [14] or to outline a set of cells that are representative of the tumor or of the specific targeted cellular regions [15], [16]. Moreover, most of these commercial tools are integrated into whole-slide scanning instruments [11], [16] or are hardware-based in that they require multispectral imaging technologies attached to microscope platforms that may be not readily available to a regular pathologist or onchologist [15], [17].

The recent literature reports many attempts to develop automated methods for nuclear segmentation. The used approaches vary among the most widely known image processing techniques, including intensity thresholding [18], [19], active contours [20], [21], [22], [23], [24], watersheds [24], [25], [26], multiscale analysis [27], a priori geometric models [28], [29], graph-cuts [30], Markov random fields [31], etc. Despite active research in this field, reliable cancerous nuclei segmentation remains extremely challenging.

The major contribution of this paper is to provide a fully automated and flexible procedure that overcomes the limitations of the existing computer-based techniques.

The paper is structured as follows: in Section 2 we discuss the main challenges related to IHC tissue image segmentation and limitations of the existing techniques and we describe the main contributions of our work; in Section 3 we present our case study related to the IHC analysis of lung cancer tissue images and we describe in detail our proposed methods; in Section 4 we discuss parameters and implementation; in Section 5 we present experimental results; in Section 6 we conclude the paper.