Page 48 - Fister jr., Iztok, and Andrej Brodnik (eds.). StuCoSReC. Proceedings of the 2018 5th Student Computer Science Research Conference. Koper: University of Primorska Press, 2018
P. 48
orporating opening-closing reconstruction and the distance The paper is organized as follows: section 2
transform with chamfer algorithm after color deconvolution, and demonstrates the problem formulation and brief implementation
H-minima. It was claimed that the proposed segmentation model of CS algorithm. Section 3 describes the experimental results and
accurately detect the nuclei and overcome the limitation of the paper is concluded in section 4.
classical watershed algorithm like over-segmentation due to the
sensitivity to noise. 2. Image Clustering using Cuckoo Search (CS)
Algorithm
Another novel approach is to segment the nuclei regions
and then resolve the overlapping or clump nuclei separation Clustering is a process of organizing data into clusters that have
through heuristic approaches like the Concave Point Detection high intra-cluster and low inter-cluster similarity. It is clear that
[11]. Wienert et. al. [11] presented novel contour-based intra-cluster similarity should be maximized and inter-cluster
"minimum-model" cell detection and segmentation approach that similarity should be minimized. Based on this idea, objective
used minimal a priori information and detects contours functions are defined [24]. The best partitioning of a given data set
independent of their shape. Experimental results proved that the can be attained by minimizing/maximizing one or more objective
proposed segmentation model capable to avoid the segmentation functions. The objective functions can be formed by capturing a
bias with respect to shape features and allows for an accurate certain statistical–mathematical relationship among the individual
segmentation with high precision (0.908) and recall (0.859). data items and the candidate set of representatives of each cluster
(also known as cluster centroids) [25].
A few researchers also presented different voting
algorithms, which cast votes along gradient directions amplifying 2.1 Problem Formulation
votes inside the centre of nuclei thereby locating the seed points
as ones having maximum votes [12, 13, 14]. Consequently, the Suppose one specific dataset consists of classes (i.e. 1,
detected nuclei seed points had been either utilized to initialize 2, … , ) and features. Therefore, the clustering problem is the
active contours [12, 13] or an edge grouping algorithm [14]. finding of the optimal position of centroids in an -dimensional
Recently, deep neural network proved its effective performance in space i.e. each centroid is an -dimensional vector. With this
breast histopathology image segmentation field [15, 16]. Su et. al. premises, the ith individual or solution of the applied optimization
[15] employed one fast Deep convolutional neural network (CNN) algorithm is a vector with . components which can be denoted
for pixel-wise region segmentation of breast histopathology
images. Experimental results proved that the proposed as follows [19, 26]:
segmentation model gave superior performance over both the LBP
feature-based and texton-based pixel-wise methods within less = (1 , 2 , … … … . ) (1)
computational time. Naylor et. al. [16] developed one hybrid Where, = (1,, 2,, … … … . ,)
nuclei segmentation model by combining deep learning and
mathematical morphology. Test results showed the promising So any solution in the population of the employed optimization
performance of the proposed model. algorithm consists of . components, each of which can take

Although, clustering based segmentation which shows any real value.
its effective performance in hematopathology [17] or other The fitness function has been calculated by summing out the
histopathology [18] image segmentation domain, but have not Euclidean distance between the data vector instance and the
been used in breast histopathology image field according to best centroid of class it belongs to according to minimum distance
of the knowledge. Therefore, this study concentrates to apply the criterion to the corresponding centroid i.e. (()) as in K-
clustering based segmentation to segmentize the different regions
of the breast histopathology images. K-means is the well-known means.
clustering techniques but sensitive to initial cluster centres and
easy convergences to local optimization. Therefore, Nature- () = ∑= 1 (, ()) (2)
Inspired Optimization Algorithms (NIOA) are successfully
employed to overcome the problems of K-means in image is the number of data vectors to be clustered. (. , . ) is the
clustering domain [19-23]. For example, Orman et. al. [20] Euclidean distance, is the ith solution of the population. So by
developed Particle Swarm Optimization (PSO) based satellite and choosing the discussed fitness function, the problem can be
MRI image clustering model and claimed that it outperformed
some state-of-the-art methods for image classifier such as K- considered as a minimization problem which is defined as
means, Fuzzy C-means, K-Harmonic means and Genetic
Algorithm based model. In this study, Cuckoo Search (CS) follows:
algorithm has been employed and compared with classical K-
means algorithms. Experimental results show that CS provides = Arg[Min∀(())] (3)
better-quality segmented images compare to classical K-means by
considering the values of objective (fitness) function, is the optimal set of centroids. In the case of image clustering,
computational time and quality parameters. depends on user, = 3 for RGB colour image, is equal to

StuCoSReC Proceedings of the 2018 5th Student Computer Science Research Conference 50
Ljubljana, Slovenia, 9 October
   43   44   45   46   47   48   49   50   51   52   53