An agglomerative segmentation framework for non-convex regions within uterine cervix images

Abstract

The National Cancer Institute has collected a large database of uterine cervix images termed “cervigrams”, for cervical cancer screening research. Tissues of interest within the cervigram, in particular the lesions, are of varying sizes and of complexnon-convex shapes. The tissues possess similar color features and their boundaries are not always clear. The main objective of the current work is to provide a segmentation framework for tissues of interest within the cervix, that can cope with these difficulties in an unsupervised manner and with a minimal number of parameters.

The proposed framework transitions from pixels to a set of small coherent regions (superpixels), which are grouped bottom-up into larger, non-convex, perceptually similar regions. The merging process is performed utilizing a new graph-cut criterion termed the normalized-mean cut (NMCut) and an agglomerative clustering framework. Superpixels similarity is computed via a locally scaled similarity measure that combines region and edge information. Segmentation quality is evaluated by measuring the overlap accuracy of the generated segments and tissues that were manually marked by medical experts.

Experiments are conducted on two sets of cervigrams and lead to the following set of observations and conclusions: 1) The generated superpixels provide an accurate decomposition of the different tissues; 2) The local scaling process improves the clustering results; 3) The influence of different graph-cut criterions on the segmentation accuracy is evaluated and the NMCut criterion is shown to provide the best results; 4) A comparison between several modifications to the agglomerative clustering process is conducted. The results are shown to be strongly influenced by the merging procedure; 5) The agglomerative clustering framework is shown to outperform a state-of-the-art spectral clustering algorithm.

Introduction

This work is focused on automatic analysis of optical images of the uterine cervix, termed “cervigrams”. Cervicography uses visual testing based on color change of cervix tissues when exposed to 3%–5% acetic acid. This helps to detect abnormal cells that turn white (acetowhite) following the application of the acetic acid [11]. In cervicography the uterine cervix is photographed with a special 35 mm camera with a ring flash, used to provide enhanced illumination of the target region. The method can be used for cervical cancer screening and permits archive and study of cervical cancer. Automated cervigram analysis tools such as cervix tissues segmentation, are needed for these tasks. An accurate delineation of the different tissues within the cervix enables the extraction of various features that can be used in successive analysis and indexing steps. The development of an automated cervix tissues segmentation framework is a challenging task, which is handled and discussed in the current work.

The images used in this work were selected from a large database of cervigrams collected by the National Cancer Institute (NCI), National Institute of Health (NIH). This database was collected as part of an ongoing effort for investigating the role of Human Papillomavirus (HPV) in the development of cervical cancer and its intraepithelial precursor lesions in women [22]. NCI in collaboration with the National Library of Medicine (NLM), NIH, is developing a unique Web-based database of the digitized cervix images to study the evolution of lesions related to cervical cancer. The current work is part of an ongoing research that targets the generation of an automated analysis framework for the cervigrams within the NIH archive.

A typical cervigram is presented in Fig. 1(a). The cervix region is the main region of interest within the cervigram. This region can be divided into three tissues of interest, as defined by NCI experts for the automatic analysis task: The squamous epithelium (SE), which is smooth and pink; the columnar epithelium (CE) that appears red and irregular; and the acetowhite (AW) region which is a transient, white-appearing epithelium following the application of acetic acid. The separation between the AW and the SE tissues presents an extremely difficult image segmentation task. The AW tissue can be identified by its color, which is lighter than the color of the surrounding SE tissue and by its boundary, but earlier studies have shown that a large overlap exists between the color distributions of these two tissues [5]. The boundary itself may vary in quality and is not always clearly visible. The AW lesions are of varying shape, size, and can be located in different places within the cervix region. Thus no shape constraints can be imposed to aid the segmentation task and the addition of position constraints is not simple. Specular reflection (SR) artifacts, generated during the image acquisition process, can be detected on the surface of the cervix. These artifacts provide misleading tissue information and interfere with the automatic segmentation.

Initial studies can be found on the analysis of individual cervigram images, or the higher-resolution colposcopic images. Part of these studies are semi-automated, requiring the user to mark regions of interest on various cervix tissues [3], [12], [20]. Features such as color [20], texture [12], and shape [3] are then extracted and used for classification of the manually extracted regions. Additional works address the task of fully-automated colposcopic image analysis [14], [21]. Works that develop novel acquisition protocols, that enforce a more controlled imaging environment or enables fusion of image modalities, are starting to emerge [1], [15], [19]. Segmentation efforts for cervigrams within the NIH database were recently introduced [2], [5], [10], [26], with most works relying on pixel-based clustering using color features [2], [5], [26].

Fig. 1(d) presents an example segmentation generated by unsupervised clustering via a mixture of Gaussians [5]. The segmentation process is performed on a preprocessed cervigram (Fig. 1(c)). Preprocessing includes the automatic detection of the cervix boundary (outlined in white in Fig. 1(b)) [28], which is done in order to focus successive analysis within the cervix region (regions outside the cervix are masked out in Fig. 1(c)). In addition, the cervix region is preprocessed for SR detection (marked in black in Fig. 1(c)) [29] and for illumination correction [5]. Each tissue is colored differently in this example, AW tissue is colored light-blue CE tissue is colored yellow and SE tissue is colored red. SR and non cervix tissues are colored dark-blue. This example illustrates the weaknesses of the pixel-based clustering framework. Due to the large overlap of tissues color distributions, only a small part of the AW tissue is detected. Furthermore, the pixel-based segmentation is very noisy and each segment is broken into many disjoint components.

The current work tries to overcome these weaknesses by adding additional information into the segmentation framework, which is mostly related to the local continuity of pixel features within the image plane. The framework shifts from a pixel-based clustering scheme to clustering of small coherent regions in the image plane termed superpixels. Such a region-based clustering scheme has the following benefits: 1) The superpixels representation augments the pixel-based color feature space to include local region and edge distributions. These distributions are statistically robust descriptors for the local image content. Using them for segmentation provides smooth segmentation maps; 2) Similar superpixels exhibit similar properties and can be grouped into larger, perceptually similar regions within the cervix. The complexity and computational cost of clustering is reduced from the number of pixels to the number of superpixels in this case.

Different region-based clustering schemes exist for segmentation of general images. Hermes et al. [9] for example, use local Mixture of Gaussians (MoG) distributions estimated from local color histograms, for the representation of rectangular regions within an image grid. The MoGs are clustered in a deterministic annealing framework, which provides segmentation results in various scales. In a bottom-up aggregation framework used by Sharon et al. [23], segment fragments of increasing size are detected and merged using their coarse scale properties. In a more recent work by O'Callaghan and Bull [18], the morphological Watershed transform is used to generate the primitive regions. The regions are then represented by histograms and clustered using a spectral clustering technique that optimizes the weighted-mean cut criterion. Another segmentation method based on spectral-clustering and the normalized-cut criterion is used by Fowlkes and Malik [6]. A multi-level region-based framework for AW tissue segmentation within colposcopy images of the uterine cervix was recently introduced by Lange [14]. This work required a substantial amount of human intervention for parameter tuning and presented initial segmentation results.

The main objective of the current work is to develop an unsupervised region-based segmentation framework for the cervigrams within the NIH database. No work has addressed this task before in a way that can cope with the large within-and across-image variability of the cervigrams with a minimal number of parameters and human intervention. The suggested framework includes the following steps: utilizing the morphological Watershed transform the image is over-segmented into a large set of superpixels (Section 2). A similarity matrix between superpixels that combines region and edge information is constructed next (Section 3). This matrix is used in an agglomerative clustering framework, which utilizes different graph-cut criteria (Section 4). The segmentation in this work is performed within the cervix region (automatically detected, or manually marked by the expert) following the preprocessing stages of SR detection and illumination correction (as shown in Fig. 1 (c)).

Specific objectives that relate to the different steps of this framework are: 1) Provide a map of superpixels that accurately decomposes the cervix tissues into small regions; 2) Compute a similarity matrix that can cope with the large variability of color and edge information that exists within and across the cervix tissues; 3) Explore the performance of an agglomerative clustering framework using different graph cut criteria, in the complicated cervigram scenario. A specific focus is placed on a new graph-cut criterion termed the normalized-mean cut that was devised for the cervigram images. This criterion enables the generation of segments that capture the elongated and non-convex nature of the tissues within the cervix region (e.g. Fig. 1(a)). The paper includes a thorough analysis of different steps within the presented framework (Section 5) and ends with a discussion (Section 6). Parts of this work were recently published [8].