A structure-aware splitting framework for separating cell clumps in biomedical images
Introduction
In medical, biological and other fields, object clumps are often observed in the captured images, which brings great difficulties to analyze the image content. For ease of analyzing, the objects are often assumed to have approximately convex shape. However, manual splitting of such convex objects tends to be time consuming and subjective [1]. As the computer vision technique is of great potential to accomplish this task automatically and accurately, developing the effective splitting framework, especially for splitting cell clumps, has gained great attention in recent years.
Essentially, many methods have been proposed to split clumps in biomedical images. The watershed method based on marker-controlled is one of the most popular methods to split clumps [2], [3], [4], [5]. However, methods based on the watershed suffer from over-splitting and have difficulties to split highly overlapped objects because the detection of markers is still not accurate. Using the variational framework of the level set is another way to split clumps [6], [7], [8], [9]. Based on the level set theory, the boundary of the objects can be obtained more accurately, but the biggest problem that restricts the application of this method is that it is computationally expensive and time consuming. Some works are based on morphology analysis [10,11]. For example, to tackle the automated morphology analysis problem of overlapping nanoparticles, [11] proposed a modified ultimate erosion process (UECS) and an edge-to-marker association to split the overlapping convex objects. Especially, the theoretical justification on the split capability of UECS was also provided.
Concave point detection is another method that is widely used in the clump splitting [1,[12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23]]. The concave point is the local curvature maximum point in the concave region formed by the clump shape. Generally, after obtaining the concave points, clump splitting could be realized by contour segmentation and ellipse fitting [1,12,13]. However, ellipse fitting is not able to split complex clumps into individual objects due to the absence of contours and the unknown number of objects in the clump. In view of this, methods in [10], [11], [12], [13], [14], [15] use concave point matching to construct split lines to achieve clump splitting. However, one limitation of these methods is that they usually pre-define some parameters to get concavity point-pairs. When the image contains objects of varying shapes and sizes, these methods may not be able to perform well. To this end, a new method using a variable-size rectangular window was presented by Farhan et al. [17] to search for the best concavity point-pair, which reduces the dependency on the user-defined parameters and increases the splitting accuracy. Besides, most of the concave point-based methods use a straight line between two points on the clump contour to split the clump. While Janssens et al. [20] proposed an alternative approach to use a spit path to achieve the accurate split of large clumps. The proposed method is robust but a little complex. A comparison of different concave point detection-based methods for overlapping convex object splitting can be referred to in [23].
From the above discussion, we can observe that the existing methods mainly have the following main problem. The first problem is that most of these methods only use the concave points as the splitting basis, which is sensitive to noise. They may not characterize the clump well when the cell clump contains multiple cells with complex structure. The second problem is that the splitting results of the existing methods still have over-splitting and under-splitting phenomenon. The reason is that these existing methods do not provide appropriate solutions to different types of clumps that may occur in practice. They can only deal with some cases like the clump that only contains two cells or more cells with relatively simple structures (e.g., the cell clumps in the later Fig. 6(a) and (b)). For the clumps with more complex or special structures (e.g., the cell clump in the later Fig. 6(d)), they often obtain under-splitting and over-splitting results.
To overcome these limitations, we propose a novel framework, which contains two carefully designed computational components as shown in Fig. 1. The first component is candidate splitting points extraction, which consists of the detection of concave points and centroids from each cell clump. In this way, the splitting basis built by our extracted clump candidate splitting points can better characterize the cell clump than the existing methods. The second component is the clump splitting. By categorizing the structures of cell clumps that may occur and using the corresponding splitting strategy, our splitting method can alleviate the under-splitting and over-splitting problem effectively.
To sum up, this work mainly has the following three-fold contributions:
- (1)
A more accurate concave point detection method and a modified centroid extraction method are proposed to obtain the useful clump candidate splitting points. Serving as the splitting basis, these candidate splitting points are important for characterizing the structure of the cell clump and implementing the splitting process.
- (2)
A simple but effective structure-aware clump splitting algorithm is proposed, where we first identify the structure of each cell clump, and then perform the splitting strategy for the corresponding structure type.
- (3)
A novel post-processing technique is proposed to deal with the cases when the cell clumps have incomplete clump structures or the concave points are missed due to poor image quality.
This paper is organized as follows. In Section 2, we present the overall framework to briefly introduce the steps contained in the proposed method. Section 3 introduces the candidate splitting points extraction process of each cell clump. In Section 4, we present our structure-aware splitting algorithm for splitting various types of clumps that may exist in the biomedical image. Experimental comparisons and discussions are presented in Section 5. Finally, Section 6 concludes the paper.
Section snippets
Overall framework description
This proposed method mainly contains two parts: candidate splitting points extraction and structure-aware clump splitting. The steps are displayed in Fig. 1.
The purpose of candidate splitting points extraction is to obtain all concave points and centroids from the input image, so that the subsequent processes for concave point matching and split line construction can be implemented based on them. To be specific, an image pre-processing step is first used to obtain the contour information of the
Image pre-processing
Given an input biomedical image, the pre-processing step starts with building the binary image. Specifically, we first binarize the input image by the Otsu's method [24]. Then the morphological open operation is used to smooth the object contours. Based on our observation, some of the contours obtained by the morphological open operation are still rough and may have small-scale fluctuations, which would influence the concave point detection accuracy. To this end, an edge-preserving smoothing
Design of the splitting algorithm
Due to the fact that there are various types of clumps in the images of convex objects, how to match the detected concave points becomes the most difficult part in clump splitting. In this paper, we define four types of clump structures based on the extracted concave points and centroids. Denote Nconcave and Ncentroid as the number of concave points and centroids, respectively. Then, the types of clump structures are defined as:
Simple structure: We define the simple structure where there are
Experimental results and discussion
To verify the proposed method, three datasets are constructed from the BBBC005 image set [35], which is downloaded from https://data.broadinstitute.org/bbbc/image_sets.html. The original BBBC005 image set contains about 20,000 cell images and there are no ground truths for the splitting results. We thus need to manually determine the ground truth for each test image. To alleviate the annotation burden, we just randomly select 90 images to test the validity of our proposed method. The 90 images
Conclusion
This paper has proposed a novel splitting framework for separating overlapping or torching cells in biomedical images. There are two parts contained in the proposed splitting framework: candidate splitting points extraction and structure-aware splitting. The former is used to characterize the contained cell clumps, while the latter is used to identify the clump structure and then construct the split line by using the corresponding splitting strategy. To further improve the robustness of our
Declaration of Competing Interest
None.
Acknowledgement
This work is supported by the National Natural Science Foundation of China under Grant no. 61773301 and 61876140, the China Postdoctoral Support Scheme for Innovative Talents under Grant no. BX20180236 and the Science Foundation of Science and Technology on Complex System Control and Intelligent Agent Cooperative Laboratory under Grant No. 181101. The original images in Dataset 1, Dataset 2 and Dataset 3, which are all downloaded from https://data.broadinstitute.org/bbbc/index.html, are
References (40)
- et al.
Automated segmentation of overlapped nuclei using concave point detection and segment grouping
Pattern Recognit.
(2017) - et al.
Automated cell nuclei segmentation for breast fine needle aspiration cytology
Signal Process.
(2013) - et al.
Automatic segmentation of granular objects in images: combining local density clustering and gradient-barrier watershed
Pattern Recognit.
(2014) - et al.
Segmentation of cell nuclei in fluorescence microscopy images: an integrated framework using level set segmentation and touching-cell splitting
Pattern Recognit.
(2016) - et al.
Morphological multiscale decomposition of connected regions with emphasis on cell clusters
Comput. Vis. Image Underst.
(2009) - et al.
Splitting touching cells based on concave points and ellipse fitting
Pattern Recognit.
(2009) - et al.
Automatic cell nuclei segmentation and classification of breast cancer histopathology images
Signal Process.
(2016) - et al.
A novel method for splitting clumps of convex objects incorporating image intensity and using rectangular window-based concavity point-pair search
Pattern Recognit.
(2013) - et al.
CHARISMA: an integrated approach to automatic H&E-stained skeletal muscle cell segmentation using supervised learning and novel robust clump splitting
Med. Image Anal.
(2013) - et al.
Clump splitting via bottleneck detection and shape classification
Pattern Recognit.
(2012)
Salient object detection employing a local tree-structured low-rank representation and foreground consistency
Pattern Recognit.
Segmentation of clustered nuclei with shape markers and marking function
IEEE Trans. Biomed. Eng.
Segmenting clustered nuclei using H-minima transform-based marker extraction and contour parameterization
IEEE Trans. Biomed. Eng.
Level set segmentation of cellular images based on topological dependence
Automated nucleus and cytoplasm segmentation of overlapping cervical cells
An integrated region-, boundary-, shape-based active contour for multiple object overlap resolution in histological imagery
IEEE Trans. Med. Imaging
Segmentation, inference, and classification of partially overlapping nanoparticles
IEEE Trans. Pattern Anal. Mach. Intell.
Segmentation of partially overlapping nanoparticles using concave points
A new separation algorithm for overlapping blood cells using shape analysis
Int. J. Pattern Recognit. Artif. Intell.
Detection and segmentation of cell nuclei in virtual microscopy images: a minimum-model approach
Sci. Rep.
Cited by (9)
Position-based anchor optimization for point supervised dense nuclei detection
2024, Neural NetworksImproving concave point detection to better segment overlapped objects in images
2024, Multimedia Tools and Applications