An automated image analysis methodology for classifying megakaryocytes in chronic myeloproliferative disorders

Abstract

This work describes an automatic method for discrimination in microphotographs between normal and pathological human megakaryocytes and between two kinds of disorders of these cells. A segmentation procedure has been developed, mainly based on mathematical morphology and wavelet transform, to isolate the cells. The features of each megakaryocyte (e.g. area, perimeter and tortuosity of the cell and its nucleus, and shape complexity via elliptic Fourier transform) are used by a regression tree procedure applied twice: the first time to find the set of normal megakaryocytes and the second to distinguish between the pathologies. The output of our classifier has been compared to the interpretation provided by the pathologists and the results show that 98.4% and 97.1% of normal and pathological cells, respectively, have testified an excellent classification. This study proposes a useful aid in supporting the specialist in the classification of megakaryocyte disorders.

Introduction

Megakaryocytes constitute a population of bone marrow resident cells which are responsible for the production of blood platelets. Among bone marrow cells, the mature megakaryocytes have a peculiar morphology, being both large, as well as having a characteristically shape nucleus (see Fig. 1a). In several pathological conditions, either reactive or neoplastic affecting the bone marrow, the normal morphology of the megakaryocytes can be altered (see Fig. 1b). In particular, in the context of Philadelphia-negative chronic myeloproliferative diseases, a group of neoplastic clonal disorders of the bone marrow stem cell comprising chronic idiopathic myelofibrosis (IMF), policytaemia rubra vera (PV) and essential thrombocythaemia (ET), megakaryocytes have been regarded as key cells on which the histopathological diagnosis is mostly based (Michiels and Thiele, 2002, Thiele et al., 2000, Thiele et al., 2001). Actually, an array of histopathological features, including the size and shape of the megakaryocyte cytoplasm and the morphology of the nucleus, has been proposed and can provide enough information, allowing a correct differential diagnosis among different chronic myeloproliferative disorders and reactive pathological conditions (Florena et al., 2004, Gianelli et al., 2006). However, the issue of an effective standardization of such features has been widely debated and still represents a matter of concern (Thiele et al., 2005).

With respect to the previous work by the same authors (Tripodo et al., 2006), this study is mainly focused on methodological problems. We aimed at developing a system which could provide a non-parametric morphological analysis of human megakaryocytes which could assist the pathologist in differentiating normal from pathological megakaryocytes and in discriminating among different myeloproliferative conditions.

The detection of components embedded in histological images is difficult and in many cases their correct classification is very hard due to, for example, closely clustered cells, color variations of the marker and aberrations caused by the optical system of the acquisition device. Beksaç et al. (1997) proposed a semiautomated approach based on a set of features; however some problems need to be solved in the case of overlapping and touching cells. A number of supervised methods allowing the relating of objects together with their characterizing features, but this approach is time-consuming, subjective and error-prone. Coelho et al. (2002) introduced a methodology based both on binary images of retinal cells and on manual segmentation.

Previous contributions on the automatic analysis of cells required a hand-made segmentation. A system to automatically segment and track cells in microscopy sequences has been presented (Yang et al., 2006), but it deals with fluorescent images that do not contain enough details for a reliable classification.

We aimed at describing a non-parametric method to find the best set of features that identify an object which is useful to the problems of classification and query in pictorial databases.

This paper reports on a methodology based on standard and well know algorithms including mathematical morphology (Section 3), wavelet analysis (Section 4) and elliptic Fourier descriptors (Section 5). In spite of the selected algorithms needing some parameters, the proposed methodology has been designed so that the user does not define any parameters. The chosen components can be extracted with a number of algorithms and many their combinations can be arranged to improve the performances. A system for image analysis could give good and robust algorithms and it could be a sufficient solution to the problem. It does not require a user to make his own choice or to find his own efficient combination of algorithms. Moreover, many algorithms are equipped with a set of input parameters that are not easy to define for an inexperienced user. To avoid these problems, our methodology introduces one of the robust sequences of the necessary algorithms to identify the pre-selected components without parameters. Such a transformation was not easy and in some cases it was necessary to include substantial variations of the original technique: an exhaustive study was necessary to normalize the data during a pre-processing step; a collection of parameters was obtained with an auto-calibration procedure in the segmentation phase. Most operations can be performed quickly to find structures with a particular shape, size and orientation (Serra, 1982, Soille, 2003). The use of watershed clustering (which is closely connected to mathematical morphology) has been proposed by Jiang et al. (2006) to locate white blood cells and by Isitor and Thorne (2007) to segment nuclei.

Different steps are required in order to discriminate the megakaryocytes (see Fig. 2). A first phase segments the image to isolate the cytoplasm and its nucleus in a given microphotograph (Section 2). This information is used in the subsequent enhancement of the shape of the cell (Sections 3 Segmentation using mathematical morphology, 4 Segmentation using wavelets) and extraction of the features which return its distinctive signature (Sections 5 Shape analysis using elliptic Fourier descriptors, 6 Feature extraction). The actual classifier is a regression tree procedure applied on the set of these signatures (Section 7). Final conclusions are given in Section 8.