Circular shape constrained fuzzy clustering (CiscFC) for nucleus segmentation in Pap smear images

https://doi.org/10.1016/j.compbiomed.2017.04.008Get rights and content

Highlights

  • A novel framework is proposed to segment nucleus from overlapping Pap smear images.

  • A new method is presented to incorporate circular shape constrain in FCM clustering.

  • ISBI 2014 challenge dataset of 900 overlapping cervical images used for evaluation.

  • The proposed method can segment nucleus with high accuracy, precision and recall.

Abstract

Accurate detection and segmentation of cell nucleus is the precursor step towards computer aided analysis of Pap smear images. This is a challenging and complex task due to degree of overlap, inconsistent staining and poor contrast. In this paper, a novel nucleus segmentation method is proposed by incorporating a circular shape function in fuzzy clustering. The proposed method was evaluated quantitatively and qualitatively using the Overlapping Cervical Cytology Image Segmentation Challenge - ISBI 2014 challenge dataset comprised of 945 overlapping Pap smear images. It achieved superior performance in terms of Dice similarity coefficient of 0.938, pixel-based recall 0.939 and object based precision 0.968. The results were compared with the standard fuzzy c-means (FCM) clustering, ISBI 2014 challenge submissions and recent state-of-the-art methods. The outcome shows that the new approach can produce more accurate nucleus boundaries while keeping high level of precision and recall.

Introduction

According to American Cancer Society (ACS) and World Cancer Research Fund, cervical cancer is the fourth most common cancer in women and 14.1 million new cases has been diagnosed worldwide in 2012. Cervical cancer arises from cervix tissues with long term infection with Human Papilloma Virus (HPV). Fortunately it can be cured at early stages. Cervix tissues undergo pre-cancerous changes over long time before true cancer cell develop, however rarely show any physical symptom. Hence, regular screening is the only way for timely detection of cervical cancer.

The Papanicolaou test (or Pap test) is the most effective screening, introduced by the Greek doctor Georges N. Papanicolaou [1], to detect abnormal changes in cervix. Fig. 1 shows examples of Pap smear images. Cytotechnologists visually examine the microscopic slides to identify abnormalities in cell morphology and structure. Visual interpretation of cervical cytology slides is a time-intensive job and requires high level of concentration from a cytotechnologist [2]. Correct interpretation of Pap smear slides faces major challenges from the quality of sampling (number of cells and overlap among the cells), smearing (presence of other elements as: mucus, blood and other debris with cervical cells), and poor contrast due to uneven staining [3]. Thus, there is a need for computer-aided diagnosis system to assist the cytotechnologists and improve the quality of Pap test outcome, which in turn decreases the cervical cancer death-rate [4].

Automatic segmentation and analysis of cervical cells has been studied for several decades by researchers to assist automated diagnostic procedures. Studies have been reported for the segmentation of cervical cell nucleus and cytoplasm [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [3], [27], [28], [29], [30], [31], [32], classification of cervical cells [3], [33] and generation of simulated Pap smear images [34]. Isolated and non-overlapping cells are the common focus in literature, however very few recent studies has taken on the challenge of overlapping cells in Pap smear images.

Precise segmentation of cell nucleus is the prerequisite and an integral part for the computer aided analysis of cervical cells and diagnostic decisions. For malignancy associated change (MAC) analysis [17], [35], accurate nucleus detection and segmentation is the essential primary step. Morphological and texture features of nucleus are the most important factors to identify normality or malignancy. As diseased, cell nucleus enlarge disproportionately, its shape may become irregular [16], or texture may change due to the change in nuclear chromatin patterns [36]. In the literature, nucleus is often used as a definite cue to locate cells. Although the segmentation and detection accuracies reported in literature are high, there are still scopes for improvement. It is evident from literature that segmentation accuracy can be improved by shape influence. This paper presents a novel circular shape constrained fuzzy clustering (CiscFC), with the aim to detect and segment nucleus more accurately from overlapping Pap smear cell images compared to the current state-of-the-art methods.

Section snippets

Related works

The most studied approaches for the segmentation of single and overlapping nuclei in cervical smear images are based on thresholding [5], [6], [7], [8], [9], [10], [11], [12], [13], [14], morphological operations [15], [16], [17], deformable models [18], [19], [20], [21], [22], watershed segmentation methods [23], [24], [25], [26], [3], and region-based methods [27], [28], [29], [30].

Thresholding is a widely used traditional approach for segmenting nucleus from cell images. In earlier attempts,

An overview

Standard FCM clustering [47] associates a data point with multiple clusters using varying degree of membership ranged from 0 to 1. Clustering process iterates to partition a set of data points, by minimizing the objective function J, with respect to the membership values of matrix U and the set of cluster centers c={cj} as:J=i=1Dj=1CUijmxicj2.Here, U={Uij} is the fuzzy partition matrix, C is the total number of clusters, xi is the ith data point in a D dimensional data-space, . is the

Implementation

The proposed framework of nucleus segmentation is composed of three stages, namely background subtraction, proposed clustering and false positive reduction. Fig. 2 illustrates the flowchart of the proposed Pap smear cell nucleus segmentation framework. In the background subtraction stage, the foreground image is obtained by removing the background from Pap image. The proposed clustering is then performed using the complemented foreground image and initialized fuzzy partition matrix. The cluster

Dataset

The proposed approach was validated on the dataset from Overlapping Cervical Cytology Image Segmentation Challenge - ISBI 2014 [30], [51]. The challenge dataset contains 945 synthetic Pap smear images with varying number of cells from 2 to 10 per image and different degrees of overlap. The challenge dataset was released over two phases. Forty five images from the first phase of the challenge, released on January 15, 2014, were used as the training dataset to determine parameters such as: number

Discussion

This study presented a novel approach to include a circular shape constrain in FCM clustering. Proposed method achieved superior Dice similarity coefficient, pixel-based recall and object based precision, and pixel-based precision and object-based recall were also comparable to the state-of-the art methods. The proposed circular shape function included in FCM clustering, enabled the method to extract nucleus more accurately, which is visible from higher pixel-based recall and DSC. DSC of 97%

Conclusion

This study proposed a framework for nucleus segmentation in overlapping Pap smear images. A novel circular shape constrained fuzzy clustering has been introduced. The proposed clustering incorporated a circular shape function to influence the partitioning and find individual nucleus more accurately from the foreground images. Experimentation was conducted using the ISBI 2014 challenge dataset of 900 images and performance of the proposed method was evaluated quantitatively and qualitatively.

Conflicts of interest

None Declared.

Ratna Saha currently pursuing Ph.D. in the field of Medical Image Analysis at Flinders University, Adelaide, Australia. She received B.Sc. in Computer Science and Engineering from Khulna University, Bangladesh in 2007. Ratna received Faculty of Science and Engineering Research Award in 2016. Her research interest includes image segmentation, image processing, medical image analysis, computer vision and machine learning.

References (53)

  • J. Matas et al.

    Robust wide-baseline stereo from maximally stable extremal regions

    Image Vision. Comput.

    (2004)
  • K.-S. Chuang et al.

    Fuzzy c-means clustering with spatial information for image segmentation

    Comput. Med. Imaging Graph.

    (2006)
  • W. Cai et al.

    Fast and robust fuzzy c-means clustering algorithms incorporating local information for image segmentation

    Pattern Recognit.

    (2007)
  • B.N. Li et al.

    Integrating spatial fuzzy clustering with level set methods for automated medical image segmentation

    Comput. Biol. Med.

    (2011)
  • S.K. Adhikari et al.

    Conditional spatial fuzzy c-means clustering algorithm for segmentation of mri images

    Appl. Soft Comput.

    (2015)
  • G.N. Papanicolaou

    A new procedure for staining vaginal smears

    Science

    (1942)
  • E. Bengtsson, Recognizing signs of malignancy-the quest for computer assisted cancer screening and diagnosis systems,...
  • R.L. Cahn et al.

    Segmentation of cervical cell images

    J. Histochem. Cytochem.

    (1977)
  • E. Bengtsson et al.

    High resolution segmentation of cervical cells

    J. Histochem. Cytochem.

    (1979)
  • H. Borst et al.

    A thresholding method for automatic cell image segmentation

    J. Histochem. Cytochem.

    (1979)
  • S.F. Yang-Mao et al.

    Edge enhancement nucleus and cytoplast contour detector of cervical smear images

    IEEE Trans. Syst., Man, Cybern., Part B (Cybern.)

    (2008)
  • A. Tareef, Y. Song, M.Z. Lee, D.D. Feng, M. Chen, W. Cai, Morphological filtering and hierarchical deformation for...
  • M.E. Plissiti, M. Vrigkas, C. Nikou, Segmentation of cell clusters in pap smear images using intensity variation...
  • Segmentation of subcellular compartments combining superpixel representation with Voronoi diagrams,...
  • H.A. Phoulady, D.B. Goldgof, L.O. Hall, P.R. Mouton, A new approach to detect and segment overlapping cells in...
  • H. Lee, J. Kim, Segmentation of overlapping cervical cells in microscopic images with superpixel partitioning and...
  • Cited by (0)

    Ratna Saha currently pursuing Ph.D. in the field of Medical Image Analysis at Flinders University, Adelaide, Australia. She received B.Sc. in Computer Science and Engineering from Khulna University, Bangladesh in 2007. Ratna received Faculty of Science and Engineering Research Award in 2016. Her research interest includes image segmentation, image processing, medical image analysis, computer vision and machine learning.

    Mariusz Bajger received the M.Sc. in Applied Mathematics degree from the Jagiellonian University in Cracow (Poland) in 1988 and the PhD degree in mathematics from the University of Queensland in 1996. Since 2002 he is a full-time lecturer with the School of Computer Science, Engineering and Mathematics at the Flinders University. His research interests include applications of mathematics and computer science to problems in medical image analysis and pattern recognition with focus on breast cancer detection in screening mammograms, whole-body CT segmentation and computational human anatomy.

    Gobert Lee received the PhD degree in Medical Image Analysis from Flinders University, Adelaide, Australia in 2004. She was nominated by the Australian Academy of Science and received a JSPS research fellowship in 2005–2006, then a research fellow in 2007–2008 at the Graduate School of Medicine, Gifu University, Gifu, Japan. Since 2009, Gobert has joined the School of Computer Science, Engineering and Mathematics as a lecturer and a member of the Medical Device Research Institute, Flinders University. Her research interests include computer-aided breast cancer screening, medical image segmentation and whole-body CT segmentation for digital human model generation.

    View full text