A review of thresholding strategies applied to human chromosome segmentation
Highlights
► Automatic segmentation and disentangling of cytogenetic images are open issues.► Images are affected by the presence of hyper- and hypo-fluorescent regions and by contrast variability.► Thresholding is often the first, crucial step in cytogenetic image analysis.► We compare eleven thresholding methods in segmenting chromosomes and chromosome clusters.► Adaptive Threshold and Region Based Level Set show the best performance overall.
Introduction
Chromosome karyotyping analysis [1] is an important screening and diagnostic procedure routinely performed in clinical and cancer cytogenetic labs. Chromosomes are first stained with a fluorescent dye, and then imaged through a microscope for subsequent analysis and classification. Each chromosome in the image has to be identified and assigned to one of 24 classes: the result is the so-called karyotype image, in which all chromosomes in a cell are graphically arranged according to an international system for cytogenetic nomenclature (ISCN) classification [2]. Fig. 1 shows four typical PAL resolution (768 × 576, 8 bits/pixel) Q-banding prometaphase images.
Individual chromosomes only appear as distinct bodies towards the end of the cell division cycle, at prophase, when they are long string-like objects, contracting and separating at metaphase, just before cell division. Most of the studies aimed at the development of automatic cytogenetics systems for the analysis of banded chromosome preparations have concentrated on the prometaphase, the intermediate stage of contraction between prophase and metaphase [3], [4].
The first step to be taken in analyzing a chromosome image is the segmentation of chromosomes and chromosome clusters from the image background. Unfortunately, the high variability in chromosome and background fluorescence intensities makes the utilization of a global threshold impractical for a satisfactory segmentation of the image, since smaller chromosomes, and tails of the chromosomes, often appear with a lighter intensity than larger ones. Moreover, due to the blurred margins of the chromosomes, to the presence of staining debris, or to the the fact that long chromosomes may touch and overlap, the first segmentation step is usually unable to identify all chromosomes as single objects, but rather presents a number of clusters.
However, the main methods used to segment cytogenetic images are still based on the evaluation of a global threshold by means of the Otsu's method [5], on a global threshold with a Local Re-Thresholding (LRT) scheme [6], [7], or on K-Means Clustering on Algebraic Moments (KM-AM) binarization [8]. In [9] a local adaptive thresholding (AdT) scheme has been proposed, but it seems to be sensitive to heavily clustered images, where the fluorescein leaking out of the chromosomes fills the region where the chromosomes are concentrated. Additionally, binary segmentation can be provided by optimizing the entropy separation between the two classes, as proposed in [10]. More recently, an adaptive segmentation scheme that globally optimizing the identification of region borders through Improved Sobel with Genetic Algorithm (IS-GA) has been proposed in [11], whereas a Multistage Adaptive Thresholding (MAT) preliminary dividing the image histogram into three segments is described in [12].
The level set technique has also proved a useful segmentation framework, especially for its ability in identifying smooth contours without topological constraint, so that the number of separate object identified and their shape may not be fixed. In order to deal with space variant contrast and luminosity heterogeneity, a Region-Based term has been introduced in the Level Set formulation (RBLS) [13], [14].
Multilevel thresholding algorithm are computationally intensive, so that hybrid optimization methods combining classical Nelder–Mead (NM) simplex or Expectation-Maximization (EM) scheme with particle swarm (PS), are proposed respectively in [15] and in [16], [17].
In the presence of so many techniques, each based on a very different rationale, there is little evidence about their relative performance when dealing with a specific problem. The aim of this work is to assess and compare the performance of all the algorithms quoted above when employed for human chromosome segmentation. We implemented the methods and then run them on a dataset of chromosome images, whose manual segmentation was performed in order to have a ground truth reference against which to compare the obtained results.
Section snippets
Chromosome data
Q-band prometaphase images are cytogenetic data obtained by staining the chromosomes with quinacrine, a fluorescent dye that concentrates in different regions of the chromosomes, giving rise to the characteristic banding patterns that identify the different chromosome types.
The images thus appear as a dark background onto which the chromosomes stand out with bright and dark banding, as shown in Fig. 1.
The dataset used in this work is composed of 37 images with PAL resolution (768 × 576 pixels, 8
Methods
In this section we briefly describe all the thresholding techniques we have taken into account. We implemented them following the description provided in the original papers, highlighting the cases in which we arbitrarily set the value of free parameters.
Table 2 summarizes with brief descriptions the underlying rationales of each method employed in this work.
Results
In order to assess the performance of the 11 methods, we run them on the dataset of 37 images and compared their output against the manual ground truth reference provided.
The performance of the various algorithms are evaluated considering both their ability to correctly identify the pixels belonging to chromosomes while rejecting pixels belonging to the background (Section 4.1), and also their ability to identify as separate objects all the blobs that have been manually segmented as separate
Discussion
The eleven thresholding methods we considered in this work can be classified in three categories: the global methods provide a threshold that is constant all over the image, the local methods provide a space-variant threshold, and the multi-threshold identify a number of grey-level intervals separated by different thresholds. Although every single method has its own peculiarities (weaknesses or strengths), the ones belonging to the same category tend to share similar values in performance
Conclusion
In this paper we compared a variety of thresholding strategies by testing them on the specific problem of segmenting chromosomes (either single or in clusters) in Q-band prometaphase images. These images present two typical problems that arise in image segmentation: appearance variability of the objects of interest throughout the image and in different images, and background inhomogeneity.
Although every single algorithm has its peculiar strong/weak points, local methods have generally better
Conflict of interest
All authors have no conflict of interest.
Acknowledgment
The authors wish to thank TesiImaging Srl for financial support and for having kindly provided chromosome images.
References (30)
Intelligent splitting in the chromosome domain
Pattern Recognition
(1989)- et al.
A new method for graylevel picture thersholding using the entropy of the histogram
Computer Graphics and Image Processing
(1985) - et al.
A multistage adaptive thresholding method
Pattern Recognition Letters
(2005) - et al.
Optimal multi-thresholding using an hybrid optimization approach
Pattern Recognition Letters
(2005) - et al.
A multi-level thresholding approach using a hybrid optimal estimation algorithm
Pattern Recognition Letters
(2007) - et al.
A hybrid simplex search and particle swarm optimization for unconstrained optimization
European Journal of Operation Research
(2007) Comparison of the predicted and observed secondary structure of t4 phage lysozyme
Biochimica et Biophysica Acta
(1975)- et al.
Automatic segmentation of metaphase cells based on global context and variant analysis
Pattern Recognition
(2008) - et al.
Automatic Karyotype Analysis
(1994) ISCN: An International System for Human Cytogenetic Nomenclature (2005)
(2005)
A modular framework for the automatic classification of chromosomes in q-band images
Computer Methods and Programs in Biomedicine
An improved classification scheme for chromosomes with missing data
Fully automatic chromosome segmentation
Cytometry
Data-driven homologue matching for chromosome identification
IEEE Transactions on Medical Imaging
Toward a completely automatic neural-network-based human chromosome analysis
IEEE Transactions on Systems, Man, and Cybernetics-Part B: Cybernetics
Cited by (54)
Cytoarchitectureal changes in hippocampal subregions of the NZB/W F1 mouse model of lupus
2023, Brain, Behavior, and Immunity - HealthDown Syndrome detection with Swin Transformer architecture
2023, Biomedical Signal Processing and ControlA novel multiphase segmentation method for interpreting the 3D mesoscopic structure of asphalt mixture using CT images
2022, Construction and Building MaterialsCitation Excerpt :This algorithm calculates a global threshold for the entire CT image based on the maximum between-class variances as a famous global threshold binarization method. Otsu’s method is advanced in its simple implementation and high computation efficiency [26,27]. However, Otsu’s method performs well when the gray-level histogram approximates a balanced bimodal type [28], which inherently requires significant differences of gray levels among components of asphalt mixture.
Three-Dimensional Imaging
2022, Microscope Image Processing, Second EditionAn extended fuzzy divergence measure-based technique for order preference by similarity to ideal solution method for renewable energy investments
2020, Renewable-Energy-Driven Future: Technologies, Modelling, Applications, Sustainability and PoliciesKEMR-Net: A Knowledge-Enhanced Mask Refinement Network for Chromosome Instance Segmentation
2024, Cybernetics and Systems