Achieving the way for automated segmentation of nuclei in cancer tissue images through morphology-based approach: A quantitative evaluation
Introduction
Bio-image processing leads to the development of diagnostic tools helping pathologists and genetists in the quantification of biological activities related to diseases as well as in the design of targeted therapies [1], [2], [3]. Immunohistochemistry (IHC) [4] is a well established imaging technique that exploits intensity of stains in tissue images to quantify the protein activity related to cancer development. Image processing techniques in this context are devoted to the accurate and objective quantification and localization of such activity in specific regions of the tissue such as cytoplasm, membranes and nuclei.
In this paper we address the problem of nuclear segmentation in immunohistochemical (IHC) tissue images, that is critical for further investigations on target protein activity [5]. The main motivations of our work are: (i) to overcome the limitations of the existing methods with a fully automated morphology-based technique; (ii) to provide a better segmentation accuracy compared to a widely used and trusted method in Computer Vision and medical image processing, namely active contours, which is not well-performing in IHC tissue image segmentation. A deep quantitative evaluation of the accuracy of our morphology-based methods and active contours is provided.
Nuclear segmentation in IHC tissue images is a challenging task due to the intrinsic complexity of tissue images and to the many sources of variability that affect IHC technique. Main challenges are related to the non-predictable size and shape non-uniformity induced by the pathological process and by the lack of homogeneity of nuclear regions both in terms of morphological and chromatic features. From the morphological viewpoint, problems for segmentation arise from the presence of overlapping cells and nuclei, which are extremely difficult to separate, as well as to the presence in the sample of other non-pathological structures (e.g. connective tissue structures, blood vessels, lymphocytes, etc.) which may lead to segmentation errors. From the chromatic viewpoint, nuclear regions are characterized by non-uniform stain intensity and color, thus preventing a trivial segmentation based on color separation. In fact, the superposition of tissue layers as well as the diffusion of the dyes on the tissue surface may bring the stains to contaminate the background or other cellular regions which are different from their specific target. Moreover, different portions of the same tissue area may be not equally enlightened and stained.
We present a fully automated nuclear segmentation technique that exploits morphological and chromatic characteristics of the tissue in order to recognize images with nuclear, cellular membrane and cytoplasm activations; this allows the method to adapt to the characteristics of the image without any user-interaction. In our technique we separate nuclei from background by exploiting local analysis of intensity distribution in the neighborhood of each cell, that depends on the expected size of the cell and then on image resolution; this minimizes the effect of noise as well as of uneven staining and inhomogeneous enlightenment; then we separate clusters through our enhanced watershed technique which merges oversplit nuclei by exploiting specific chromatic characteristics of the cells.
Our morphology-based technique is compared to active contours, that is a well known and trusted technique in biomedical image processing [6], [7], [8], [9], [10]. However in the context of IHC tissue imaging remarkable intensity variations inside and outside the cellular regions to be segmented stress the limitations of active contours, calling for morphology-based approaches where the specific features of IHC images can be effectively expressed and exploited. Extensive experimental results show that our method is more accurate than various formulations of active contours in segmenting nuclei in IHC tissue images. Besides accuracy, another key metric to evaluate a bio-image processing tool is its autonomy with respect to operator input. Active contours is generally a semi-automated method that requires the operator to define a curve which the algorithm cripples iteratively to fit the boundary of the target region. On the contrary, the morphology-based technique we present in this paper provides a completely automated nuclear segmentation. To enable fully automated procedure, we also developed a pre-classification step to distinguish images showing nuclear protein activation from membrane and cytoplasm activation, so that a pathologist can give any set of tissue images as input to the tool which will recognize the type of activation and provide the quantification result as output.
A quantification technique for detection of protein activity in cytoplasm/membranes has been already presented in our previous paper [5]. In this work, we integrate the proposed nuclei segmentation technique in a fully automated tool for protein activity quantification in cancer tissues in all cell compartments (either membrane, cytoplasm or nuclei). Compared to our previous work [11], this paper provides a new set of experimental results on which a novel statistical method is applied in order to obtain more extensive comparison of morphology-based and active contours technique. Moreover we extend the examination to all the most widely known typologies of formulations of active contours present in literature, providing for each typology exhaustive discussion of its limitations in the segmentation of IHC tissue images.
The paper is organized as follows. Section 2 explains the morphological procedure. Section 3 discusses active contours based approaches. Section 4 describes the implementation details and Section 5 shows experimental results. Section 6 concludes the paper.
Section snippets
Proposed approach: morphology-based procedure
In this section we present our fully automated morphology-based procedure for nuclear segmentation in IHC tissue images. In this procedure well known morphological operators and novel techniques which exploit morphological and chromatic characteristics of the tissue are applied in order to solve the technical problems related to the segmentation of nuclei in IHC images that were already discussed in the previous section.
The accurate tracking of nuclear membranes is fundamental in IHC analysis.
Alternate approach: active contours
Our morphology-based method is compared in this work with active contours (also called snakes), a highly popular approach in Computer Vision that was successfully used in several applications including medical imaging [24], [25]. In this section we describe how we applied active contours to IHC tissue image segmentation.
First of all, preliminary tests were carried out in order to select the most effective formulations of active contours for the segmentation of nuclei in IHC tissue images; then
Implementation
The procedures were implemented in Java as plugins for ImageJ [42], a public domain software for image analysis and processing. We inherited the whole class hierarchy of ImageJ 1.38 API and the open-source plugins and macros for color deconvolution, local thresholding and spline-based snakes [18], [43], [40] and we implemented our own functions and classes. The parameters of both the methods were set by running experiments on real IHC images. See our URL [41] and [22] for details.
Experimental results and discussion
Our morphology-based method and the active contours approach were compared on the same real IHC images, on about 800 nuclei. These images showed lung cancer tissue stained by H-DAB and were acquired through high resolution confocal microscopy with three different enlargements (200, 400 and 800). Ten different histological samples were used to take the pictures: five of these samples showed activations of the target protein in the cellular membrane of the cancerous cells, two in the cytoplasm
Conclusions
In this paper we addressed the problem of nuclear segmentation in cancer tissue images comparing the effectiveness of morphology-based technique and active contours approach. Extensive experimental results show that our fully automated morphology-based technique is more accurate in IHC tissue images than various formulation of state-of-the-art semi-automated active contours, since it is less sensitive to intensity and color variations within the target region as well as overlapped nuclei that
Acknowledgments
The authors would like to acknowledge Dr. Marco Volante and Dr. Ida Rapa of Hospital S. Luigi, Orbassano, Torino.
Santa Di Cataldo is PhD student at Politecnico di Torino. Her research concerns bioimaging and automated algorithms for systems biology.
References (45)
Using active contours and mathematical morphology tools for quantification of immunohistochemical images
Signal Process
(1998)Applying deformable templates for cell image segmentation
Pattern Recogn
(2000)A semi-automatic method for quantification and classification of erythrocytes infected with malaria parasites in microscopic images
J Biomed Inform
(2009)Fronts propagating with curvature-dependent speed: algorithms based on Hamilton–Jacobi formulations
J Comput Phys
(1988)From genomics to clinical molecular imaging
Proc IEEE
(2005)Unsupervised tissue microarray analysis for cancer research and diagnostics
IEEE T Inf Technol B
(2004)Markers of small cell lung cancer
World J Surg Oncol
(2004)Pharmacodiagnostics and targeted therapies a rational approach for individualizing medical anticancer therapy in breast cancer
Oncologist
(2007)Computer-aided evaluation of protein expression in pathological tissue images
Unsupervised segmentation based on robust estimation and color active contour models
IEEE T Inf Tech B
(2005)
Level set analysis for leukocyte detection and tracking
IEEE T Image Process
Color image segmentation for objects of interest with modified geodesic active contour method
J Math Imag Vis
Segmentation of nuclei in cancer tissue images: contrasting active contours with morphology-based approach
Quantification of diverse subcellular immunohistochemical markers with clinicobiological relevancies: validation of a new computer-assisted image analysis procedure
J Anat
Graphical methods for quantifying macromolecules through bright field imaging
Bioinformatics
Segmentation of whole cells and cell nuclei from 3-D optical microscope images using dynamic programming
IEEE Trans Med Imag
Quantification of histochemical staining by color deconvolution
Anal Quant Cytol Histol
Comparison of quantification of histochemical staining by Hue-Saturation-Intensity (HSI) transformation and color deconvolution
Appl Immunohistochem Mol Morphol
Picture thresholding using an iterative selection method
IEEE T Syst M Cy, SMC-8
Automatic segmentation of digital micrographs: a survey
Image segmentation through adaptive thresholding
Technion – Israel Inst Technol
Cited by (32)
Fast unsupervised nuclear segmentation and classification scheme for automatic allred cancer scoring in immunohistochemical breast tissue images
2018, Computer Methods and Programs in BiomedicineCitation Excerpt :We adopted this commonly employed representation procedure in descriptive statistics because it is principally helpful to demonstrate the diversity between populations without making any hypothesis of the underlying statistical distribution [29]. Each box in the Fig. 8 contain horizontal lines at the three data distribution's quartiles, i.e. the 3 values which separate the ranked data set of Allred score values into four similar groups, each group comprising a quarter of the data [36]. The bottom and top of the box are always the first and third quartiles, and the band inside the box is always the second quartile (the median score value which is marked with red color in Fig. 8).
Feature based fuzzy inference system for segmentation of low-contrast infrared ship images
2016, Applied Soft Computing JournalA new automatic image analysis method for assessing estrogen receptors' status in breast tissue specimens
2013, Computers in Biology and MedicineCitation Excerpt :Millions of people die each year due to various types of cancer diseases and breast cancer is considered as one of the main causes of death among women. It has been shown that the ER status is an important biologic factor in predicting the response to hormonal therapy of breast cancers [1–6]. In daily clinical practice, IHC staining is mostly done to quantitate expression of ER in breast tissue specimens to further classify the cancer.
Automatic image segmentation of nuclear stained breast tissue sections using color active contour model and an improved watershed method
2013, Biomedical Signal Processing and ControlCitation Excerpt :For this reason, we seek to improve the FLD-MNN nuclei classification by using the proposed geometric active contour model described in Section 2.2.2.2. Some works heavily question the use of active contours methods in IHC images, due to their sensitivity to curve initialization and the lack of convergence caused by other attractors located close to the target [9,10]. In fact, all active contour algorithms applied in these works [9,10], are fed with initial boundaries manually traced by a pathologist and a monochromatic image used for the calculation of the image energy term.
Computer-aided techniques for chromogenic immunohistochemistry: Status and directions
2012, Computers in Biology and MedicineCitation Excerpt :Widely used in nuclear segmentation methods are also edge-based techniques (e.g. Canny [76,77]) and active contours [78–84,72]; nevertheless, these approaches suffer due to remarkable intensity variations within as well as outwards nuclei. In particular, active contours performance [78–84,72] is highly impacted by the noise conditions and the massive presence of foreign particles which may deviate the initial curve from the target boundary [115]. Edge-based techniques include also multi-scale analysis [85], that can handle intensity variations but is led to errors by overlapping nuclei with weak edge information.
AR- assisted intelligent analysis and identification system for mobile vegetables diseases based on HOG- SVM
2023, Proceedings of SPIE - The International Society for Optical Engineering
Santa Di Cataldo is PhD student at Politecnico di Torino. Her research concerns bioimaging and automated algorithms for systems biology.
Elisa Ficarra is Assistant Professor at Politecnico di Torino. Her research concerns bioimaging, algorithms for gene expression analysis, gene classification, clustering and networks, clinical genomics and systems biology.
Andrea Acquaviva is Assistant Professor at Politecnico di Torino. His main interests are distributed embedded software design and wireless sensor networks. His research also includes bioimaging and systems biology.
Enrico Macii is Full Professor at Politecnico di Torino. His main interests are computer-aided design of digital integrated circuits and systems, with particular emphasis on logic synthesis, optimization, testing, and formal verification. His research also includes bioinformatics, bioimaging and systems biology.