Comparison of nuclei segmentation result with the topranked approach from ISBI cell tracking challenge

Get Complete Project Material File(s) Now! »

Cells/Nuclei segmentation

Cell segmentation arises as the process of labeling every pixel in an image and therefore pixels with the same label share certain characteristics. The ultimate goal of this process is to extract interesting regions, i.e., cells or nuclei from an image. In image analysis, segmentation is considered as a fundamental step since its impossible to get good tracking results with poor segmentation even with using powerful tracking techniques. Many scientists contributed with a lot of efforts into developing various algorithms for automated segmentation of cells. However, there are still no systems that can handle the extensive variability during time-lapse experiments of cells. Additionally, several cell segmentation methods use a preprocessing step to reduce the noise and enhance the image quality as mentioned in (Section 2.1) toward better segmentation results.
The cell nuclei segmentation approaches are broadly classified into three categories: simple approaches such as thresholding method [Bise 2015, Liao 2016, Gul- Mohammed 2014a], edge detection [Wählby 2004] and shape matching [Cicconet 2013,Türetken 2017,Türetken 2015], and more sophisticated approaches like region growing [Cliffe 2017, Liu 2014,Tonti 2015,Gul-Mohammed 2014b] and energy minimization [Tarnawski 2013] and machine learning based approaches [Ronneberger 2015,Akram 2016b, Liu 2017,Sadanandan 2017].
In the first category, Liao et al. [Liao 2016] applied a classical thresholding approach to segment cell nuclei. This approach assumed that the cell was usually brighter than its surrounding areas and there often existed an optimal threshold where individual cells can be segmented as separate objects. This assumption was not applicable for challenging regions, because it was impossible to find a suitable threshold to separate all touching cells.
Bise and Sato [Bise 2015] adapted a multiple-level thresholding technique to identify and segment all cell nuclei. In this method, a set of level thresholds that were equally spaced had been used to segment cell nuclei at a particular level of intensity with the fact that the intensities inside the cell were slightly higher than those at its boundaries among touching cells.

Cells/Nuclei tracking

Cell tracking has received great attention over the past few decades therefore many approaches have been proposed to perform this task. Reviews of existing cell tracking approaches can be found in [Rohr 2010,Kanade 2011,Meijering 2012]. Besides, the general description of the approaches implemented in cell tracking challenges have been reviewed in [Maška 2014, Chenouard 2014, Ulman 2017]. Broadly speaking, cell tracking approaches can be categorized into two groups: (1) model-based contour evolution approaches and (2) detection-based association approaches.
The model-based contour evolution approaches are getting started by segmenting the cells contours in the first frame of a video sequence and then evolving these contours dynamically to fit the cells in successive frames. Thus, both segmentation and tracking tasks can be solved simultaneously as a one process, with the hypothesis that obvious spatiotemporal overlapping exist between the corresponding cell regions. These approaches represent cells contours using either a parametric or an implicit model. In the former model, active contours [Dufour 2005], active meshes [Dufour 2011], or Gaussian mixture [Amat 2014] are employed to represent the cells, which require more steps to handle key events such as cell division, cells touching and cells entering/leaving the field of view. In the latter model, the contours of the cells are represented in an implicit manner through the level-set method. Such an approach can handle changes of topology and provide enough support to the estimation of the geometric properties, but it is computationally expensive. The main advantage of contourbased evolution approaches is that they can provide robust segmentation of cell boundary even in a noisy environment. Moreover, such techniques are very useful in tracking various cell topological changes that occur during the cell cycle and growth.

Images with sparse representation.

After obtaining the optimal dictionary and sparse representation coefficients, the denoised image and detection map are reconstructed as explained in Algorithm 3, which will be used later in the detection and segmentation of cell nuclei. The denoised image with sparse representation is formed by solving the following optimization problem: ˆX = argmin X kX − Y k22 + kX − DAk2.

Cell nuclei segmentation.

In time-lapse fluorescence microscopy images, nuclei appear as bright objects on a dark background. Consequently, nuclei can be neatly extracted from the background using a simple thresholding method, but such method unable to segment touching nuclei.
In previous research, a watershed-based method was proven to be an effective tool for segmenting such touching nuclei scenarios. However, classical watershed approach usually leads to an over-segmentation problem. In this section, we present how a thresholding-based approach is used to get initial cell nuclei segmentation. Then, we discuss the generic marker extraction method, which based upon both the denoised image and detection map (Section 3.1.2, Section 3.1.2). Afterward, we briefly review the watershed-based algorithm and the marker-controlled watershed. Algorithm 4 present the main steps of cell nuclei segmentation process.

READ  Numerical implementation of beta-PDF integration  Study of convergence of the semi-analytic beta integration method

Initial cell nuclei segmentation.

A local adaptive-thresholding approach [Bradley 2007] is applied to the denoised image, in which the general concept of the algorithm is that for every image’s voxel the threshold is determined by the following equation: Tlocal = meanlocal (1 − SensitivityFactor) (3.12).
where, meanlocal is the mean intensity value in the neighbourhood of each voxel and the SensitivityFactor is a scalar value within a range from zero to one which controls sensitivity towards thresholding more voxels as foreground. Accordingly, voxels with intensity values larger than Tlocal are set to 1, all others are set to 0. Small regions detected as foreground and smaller than a predefined volume denoted by MinNucleiV olume are discarded. This threshold corresponds to the volume of the smallest cell nucleus and is determined prior to the segmentation step. The resulting image is called the segmentation mask. Alternatively, the global Otsu’s thresholding-method [Otsu 1979] can also be employed to obtain the segmentation mask, in which the basic idea of Ostu method is to iterate through all the possible threshold values and to choose the optimal one that minimizes the within-class variance. The within-class variance is simply defined as a weighted sum of variances of the two classes: 2w (t) = w2 b (t)2 b (t) + w2 f (t)2 f (t).

Marker points detection.

For splitting of touching cell nuclei, we employed a marker-controlled watershed technique. The marker points are obtained as follow: first, the denoised image is multiplied by the detection map to provide a maximum response image. Second, The maximum response image is processed to detect the local maxima (Figure 3.4). The obtained local maxima image is multiplied by the segmentation mask to discard local maxima detected in the background. Third, a morphological dilation operator of certain radius denoted by NucleiSeedDilation is employed to avoid detection of multiple local maxima for the same object by merging those maxima that were in close proximity to each other. Finally, the modified image determining the marker points is fed to the subsequent watershed algorithm.

Table of contents :

1 Introduction 
1.1 Biological Background
1.1.1 Cell cycle
1.1.2 Model organisms for studying the cell cycle
1.2 Microscopy imaging techniques
1.2.1 Laser scanning confocal microscopy (LSCM)
1.2.2 Spinning disk confocal microscopy (SDCM)
1.2.3 Light sheet fluorescence microscopy (LSFM)
1.3 Noise in fluorescence microscopy images
1.3.1 Photon noise
1.3.2 Dark noise
1.3.3 Read noise
1.3.4 Sample noise
1.4 Image analysis
1.5 Cell tracking challenge
1.6 Problem statement
1.7 Research objectives
1.8 Structure of the thesis
2 Literature Review 
2.1 Denoising of microscopy images
2.2 Cells/Nuclei segmentation
2.3 Cells/Nuclei tracking
2.4 Existing software
2.4.1 ImageJ/Fiji
2.4.2 Icy
2.4.3 CellProfiler
2.4.4 ilastik
2.4.5 Imaris
2.4.6 Amira
2.5 Summary
3 Methodology 
3.1 Denoising of 3D cell nuclei images
3.1.1 An introduction to sparse representation
3.1.2 Images with sparse representation
3.2 Cell nuclei segmentation
3.2.1 Initial cell nuclei segmentation
3.2.2 Marker points detection
3.2.3 Watershed and marker-controlled watershed segmentation.
3.3 Evaluation method and metrics
3.4 Implementation details
4 Results and Discussion 
4.1 Datasets description
4.1.1 Synthetic dataset
4.1.2 Real dataset
4.2 Experimental setup and suitable parameters selection
4.3 Results of denoising 3D cell nuclei images
4.4 Results of segmentation 3D cell nuclei images
4.4.1 Comparison of nuclei segmentation result with the topranked approach from ISBI cell tracking challenge
4.4.2 Comparison of nuclei segmentation result with deep learningbased method
4.4.3 Segmentation results over different 3D biological datasets.
5 Conclusion and Future Directions 
5.1 Conclusion
5.2 Future Directions
A Appendix Example 


Related Posts