Automatic Nuclei Detection in Histopathological Images based on

Convolutional Neural Networks

Roaa Saﬁ Abed Alah

, Gokhan Bilgin

1,2 a

and Abdulkadir Albayrak

1,2 b

Dpt. of Computer Engineering, Yildiz Technical University, 34220 Istanbul, Turkey

Signal and Image Processing Laboratory (SIMPLAB), Yildiz Technical University, 34220 Istanbul, Turkey

Keywords:

Nuclei Detection, Histopathological Images, FCM Algorithm, Convolution Neural Networks, Deep Learning.

Abstract:

Analysis of cells in histopathological images with conventional manual methods is relatively expensive and

time-consuming work for pathologists. Recently, computer aided and facilitated researches for the diagnostic

algorithms have obtained a high signiﬁcance to assist the pathologists to extract cellular structures. In this pa-

per, we are compering the conventional fuzzy c-means (FCM) clustering method with the proposed automated

detection system based on Tiny-Convolutional Neural Network (Tiny-CNN) to detect center of nucleus in

histopathological images, Also, in this study, we are tried to ﬁnd center of nucleus by combined unsupervised

method (FCM) with supervised method (Tiny-CNN). Brieﬂy, First step, nuclei centers are detected with FCM

algorithm which is applied as a clustering-segmentation method to perform segmentation of nucleus cellular

and nucleus non-cellular structure to ﬁnd the correct center of nuclei. Second step, the deep learning method

is used to detect center of nucleus based automated method. Afterward, combined each of these individual

methods to evaluate our model for extracting the center of nucleus on two different data set the University of

California Santa Barbara’s UCSB-58 data set and data set University of Warwick’s CRC-100 data set.

1 INTRODUCTION

Histopathological image analysis has various chal-

lenges, especially in nucleus detection. The devel-

opments in techniques that have been occurred in this

area can be a valuable assistance to the accurate di-

agnosis pathologically. However, it is still harder to

reach more details without solving the main compli-

cated issues in cellular structures especially the clut-

tering problem in tissue (Moita et al., 2018). The

cellular structure analysis is signiﬁcant in patholog-

ical diagnoses of breast cancer which depends on cell

analysis separately (Albayrak and Bilgin, 2016).

Traditionally, seed detection is a classical problem

of computer vision for the conventional of histopatho-

logical images. New researches have implemented

conventional methods various ways of deep learning

scenarios, and scenarios connecting both methods. A

graph-cut methods took the place by multi-scale ﬁl-

tering (LoG), adaptive scale selection, and a second

graph-cut operation (Al-Kofahi et al., 2010). Nu-

cleus detection in recent year, using of pre-processing

https://orcid.org/0000-0002-5532-477X

https://orcid.org/0000-0002-0738-871X

morphological methods assists to smooth and re-

move non-nuclei objects from input images. Gen-

eralized LoG ﬁlters and mean-shift clustering were

employed to detect nuclei seeds centers and false

nucleus were removed by adaptive thresholding (Xu

et al., 2017). Apply of generalized LoG (gLoG) ﬁlters

to develop correctly guess of nucleus shape detector

performance and used watershed segmentation tech-

nique to separate nucleus and count cells (Kong et al.,

2013).The domain of FCM for detecting nucleus posi-

tions of feature similarity index measure FSIM is used

as seeds for segmentation,this method is started with

FCM clustering and then using FSIM based template

matching approach for nucleus detection. FCM in

probabilistic model and a derivation of an algorithm

for FCM clustering which is supplied expanded abil-

ity through the conventional FCM (John et al., 2016).

This study which is known as Bayesian Fuzzy Clus-

tering that has extended each of variable number of

clusters and a particle ﬁlters dedicated technique to

evaluate the model parameter included the number of

clusters (Glenn et al., 2015). Detection of nuclei in

an automated methods of histopathological images is

represented as a good method. Especially, in recent

year with development in machine learning methods

Alah, R., Bilgin, G. and Albayrak, A.

Automatic Nuclei Detection in Histopathological Images based on Convolutional Neural Networks.

DOI: 10.5220/0007484301930200

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 193-200

ISBN: 978-989-758-353-7

193

which is displayed an effective way of detecting a cell

in an image by predicting for each pixel location. One

of these methods has discovered the location of each

pixel that is the monotonous function of the distance

to evaluate the closest cell to the center. Then the cell

center can be speciﬁed by extracting local extremism

of the predicted values (Kainz et al., 2015). While

Stacked Sparse Autoencoder (SSAE) method is pre-

sented to solve the complexity of automated nuclei

detection. A deep learning method is applied to detect

an effective nuclei on high-resolution histopathologi-

cal images of breast cancer which is learning high-

level features in just one-pixel intensity to recognize

distinctive features of nucleus with a good results

(Xu et al., 2016). Authors have proposed a method

based on convolution neural networks for detection

of the nuclei of cells in images with overlapped cells.

CNN is trained from all patches extracted from the

training images, which the center of pixel in training

patch are classiﬁed to three classes, background, cy-

toplasm, and nucleus. To make network learn how to

guess right classify of each central pixel of the patches

(Braz and Lotufo, 2017). This method that applied the

colour deconvolution to reconstruct all of the applied

stains. In addition, the structure of a large feature

set and modiﬁcation AdaBoost to create two detectors

that focused on various features in the appearance of

nuclei and modiﬁed of AdaBoost, is able to calculate

the cost of each feature during the selection of nuclei.

The output of each detector is incorporated by the op-

timal active contour algorithm to smooth the border

of detected nuclei (Vink et al., 2013).

A convolution neural network with the LoG ﬁl-

ter is applied to histopathological images for per-

forming nuclei detection by sliding window to whole

image.Application of this method in histopathoogi-

cal images performs promising results in nuclear de-

tection (Khoshdeli et al., 2017). The CNN based

deep learning algorithm with superpixel analysis has

shown improved segmentation performance for nu-

clei segmentation while comparing with the state-of-

the art methods. This method, over-segmenting the

original image by generating superpixels which al-

low the CNN to learn the localized features better in

the training stage. In another study, Spatially Con-

strained Convolutional Neural Network (SC-CNN)

which forced spatial constraint at the prediction of

the likelihood of a pixel by assigning higher proba-

bility values to the pixels located in the vicinity of the

nuclei centers is implemented for nucleus detection.

In that study the classiﬁcation of nuclei is presented a

Neighboring Ensemble Predictor method (NEP) com-

bined with convolution neural network. That provides

a good availability in the results of joint detection and

classiﬁcation of the proposed SC-CNN and NEP (Sor-

napudi et al., 2018),(Sirinukunwattana et al., 2016).

Automatic cell segmentation in histopathological im-

ages via two-staged superpixel-based algorithms has

been studied in (Albayrak and Bilgin, 2018).

The goal in our study is to detect the center of

nuclei on histopathological images, using FCM clus-

tering which is extracted the foreground from back-

ground cellular structure of histopathological images.

FCM obtained a high accuracy in clustering in order

to increase the ability of detecting center of nuclei in

conventional method with automated center detection

by using convolutional neural networks.

The organization of the paper is as follows: In sec-

tion 2 we describe the methodology of the study. Sec-

tion 3 describes the materials used, Nucleus detection

and technical details. Section 4 refers to results and

discussion, and Section 5 concludes the paper.

2 MATERIAL AND METHODS

In medical image technology had become possible

to obtain more meaningful information by applying

pre-processing of histopathological images. In recent

years, the development which happens in deep learn-

ing architectures, had increased quality of data analy-

sis to obtain more information of related image.

2.1 Fuzzy C-means Algorithm

Clustering method is a method of locating a set of ob-

jects into group. In this instance, the clustering algo-

rithms can be classiﬁed into two classes classes: hard

clustering and soft (fuzzy) clustering. One is hard

clustering; second one is soft clustering.In this paper,

a soft clustering method, fuzzy c-means, is applied to

obtain cellular structures in histopathological images.

Soft clustering FCM in many cases more resilient than

other hard clustering method. In soft clustering, data

elements belong to more than one cluster, and associ-

ated with each element is set of membership levels as

mentioned in (Suganya and Shanthi, 2012). In the na-

ture work of FCM algorithm is depend on assigning

membership for each class which is based on degrees

among zero and one, to indicate their partial mem-

bership. The clustering centers calculated in FCM is

shown in Eq. 1:

∑

j=1

i j

∑

j=1

i j

(1)

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

194

where, j stands for each cluster center, x

is the objects

in j and u

i j

is the degree of membership of x

, the

value u

i j

is the degree of membership of x

in cluster

j and, m is representing any real number bigger than

The membership matrix of u

i j

is calculated by us-

ing x

and c

as follows:

i j

∑

k=1



i j

k j



2/m−1

(2)

The FCM algorithm works by choosing the mem-

bership to each of data point identical to each cluster

center depending on the distance between the cluster

center and the data point the data which is closer to

the cluster center(Haﬁane et al., 2008).

2.2 Convolutional Neural Networks

(CNN)

Convolutinal neural network is a deep learning based

algorithm proposed by Lecun et al. to classify hand-

writing digit images (LeCun et al., 2010). CNN is

consisted of three main types of layers: convolu-

tional layer, pooling layer and output layer. Convo-

lution layers,pooling layer, and an output layer. Net-

work layers are compatible in a feed-forward, back-

propagation neural networks construction: each con-

volution layer is followed by a pooling layer, and the

last convolution layer is followed by the output layer.

The convolution and pooling layers are represented as

two-dimensions layers, while the output layer is con-

sidered as a 1-D layer. In convolutional neural net-

work, each 2 dimension layer has several levels. In

one level it can observe neurons that are arranged in

a 2-D array. Also, with respective of the output layer

we can obtain a feature map.

In a convolutional layer, each level is linked to one

or more feature maps of the next layer. This con-

nection is adjust together with a convolution mask,

that is a two-dimensional matrix of adaptable entries

weights term. All levels computes the convolution

among its two-dimensional inputs and its convolu-

tion masks. The convolution outputs are collected and

added weight and bias. Then, an activation function

is utilized in the result to gain the output which is a

2-D matrix recognized as a feature map. A convolu-

tion layer output generate one or more feature maps.

All feature map is connected to exactly one level in

the output of the convolution feature map as shown in

Eq. 3:

= f

(

∑

i∈v

(l−1)

⊗ w

i,n

+ b

) (3)

where l stands for the number of layers in the lay-

ers numbers in the neurons of network f

deﬁnes as

activation function, k

is a feature map output with

size H

(l−1)

×W

(l−1)

pixels, size of mask convolution

mask in this equation H

(l−1)−r

×W

(l−1)−c

, with

⊗ as 2D convolution operator v

represents list of lev-

els in layers l-1. Pooling layer is applied to decrease

the size of the feature map for further analysis in the

next layers. As an example, the sum of four pixels

is calculated and multiplied with weight before it is

added to bias term. Then the result is sent to the ac-

tivation function to obtains an output for the 2 × 2

block. At the end, each pooling level minimize its

input size into half, along each dimension. A feature

map in this layer is connected to one or more level,

then it must be connected with next convolution layer

feature map for this layer obtained by Eq. 4:

= f

(l−1)

× w

+ b

) (4)

where, the feature map n is divided into blocks each of

them consist of 4 pixels (2 × 2) in layer (l-1), where

(l−1)

is a matrix of that result concluded from the

four pixels in (l − 1). In the last convolution layer,

each level in this layer connected to one of feature

map and convolution masks are used that have the ex-

act same size as its input feature maps as proposed

in (Hatipoglu and Bilgin, 2017; Phung and Bouzer-

doum, 2009). In this study, the center of nuclei detec-

tion is proposed in Tiny-CNN and FCM with com-

bined Tiny-CNN, that we obtained a successful re-

sult of automated methods with two histopathological

data sets of UCSB-58 and CRC-100.

3 EXPERIMENTAL STUDIES

3.1 Histopathological Image Data Sets

The ﬁrst experimental data set (UCSB-58) is ob-

tained from David Rimm Laboratory in Department

of Pathology at Yale University of California, Santa

Barbara. The data set is selected from a part of UCSB

biosegmentation benchmark and includes 58 set of

histopathological images 200 × 200 size of pixels

consist of 32 benign and 26 malignant tissue images

(Gelasca et al., 2009). The size of pixel in this images

is 896×768. Additionally, each of this UCSB-58 data

set images is corrupted labeled by pathologists to be-

come 200 × 200 pixels images with ground truth as

shown in Fig. 1.

The second data set (CRC-100) is obtained from

Department of Computer Science at University of

Warwick (Sirinukunwattana et al., 2016). This study

Automatic Nuclei Detection in Histopathological Images based on Convolutional Neural Networks

195

Figure 1: A benign image sample from UCSB-58

histopathological image data set (a) and its ground-truth (b);

a malignant image sample (c) and its ground-truth (d).

Figure 2: A CRC-100 histopathological image data set sam-

ple (a) and its ground-truth (b).

involves all the 10 E&H stained histopathological

slide images in size of 500 × 500 at a resolution of

0.55µm/pixel (20×optical magniﬁcation). For detec-

tion purposes 29,756 nuclei are marked at the center

as shown in Fig. 2 and out of 22,444 nuclei centers

have an associated class label for classiﬁcation pur-

pose.

In this paper, we examine FCM clustering re-

sult with the results of tiny convolutional neural net-

works on histopathological images. The concept of

this study is consisted of two stages: First, using

the clustering method to extract the object from the

histopathological images and detecting the center of

the object, Second, the detection center of nucleus lo-

cation that is tried to obtain by Tiny-CNN processing.

Third, combined FCM results with Tıny-CNN results.

3.2 Nucleus Detection by using Fuzzy

C-Means Clustering Method in

Histopathological Images

In this paper, FCM segmentation algorithm is used to

detect to detect the center of nucleus of histopatholog-

ical images. The FCM method is used with two ap-

proaches, the ﬁrst approach is to detect the center of

Figure 3: Regional outputs of ﬁve different clusters ob-

tained by FCM algorithm.

Figure 4: a) Five different cluster regions of FCM algorithm

b) obtained nuclei cluster of FCM approach and c) ﬁnal out-

put nuclei centers image.

nucleus, and the second approach is to segmentation

of the nuclei in histopathological images, with mask

image for extracting a right nucleus of histopatho-

logical images. The FCM semiautomatic approach

which means, that we select the closest cluster man-

ually, 5 clusters are selected, since the clusters repeat

themselves after ﬁve number of clusters as shown in

Fig. 3. Then the standard deviation of these clus-

ters are calculated, the maximum standard deviation

of these clusters are selected. Afterwards, the clos-

est cluster for the ground-truth accompanied with the

data set is selected manually, as shown in Fig. 5c. Af-

ter ﬁnding the closest cluster we apply a morpholog-

ical operation on the FCM cluster result as shown in

Fig. 5d. The FCM automatic mean not chosen manu-

ally as shown in Fig. 4b to detect the center of cells as

annotated in Fig. 4c, we have experimented this ways

which is summarized, ﬁnding the maximum Standard

deviation of ﬁve clusters which are sorted to ﬁnd the

maximum cluster from the candidate clusters to use it

for detecting the center of nucleus in histopathologi-

cal images.

The segmentation steps consist two stages:

Firstly,the otsu based thresholding method is applied

to each image for extracting the regions of cellular

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

196

Figure 5: a) An original image from UCSB-58 data set, b)

its ground-truth, c) FCM result after choosing closest clus-

ter manually, d) FCM result after morphological operation,

e) FCM nuclei center detection and f) masked image of nu-

cleus with center detection.

structures. Then a dilation operation is applied to the

result of Otsu’s method. After that the binary image

obtained from the otsu’s method is multiplied to the

original image to eliminate the posibly non cellular

structures. Binary dilation of A by B, indicated to

A ⊗ B which is deﬁned as:

A ⊗ B = z|

(B) z ∩ A 6= Ø (5)

In this equation,

(B) represents a reﬂection of the

structuring element B . In other words, it is the set

of pixel locations z, the reﬂection structuring element

overlaps with foreground pixels in A when translated

to z.

3.3 Nuclei Detection by using

Tiny- Convolutional Neural

Networks

In our model the tiny convolutional neural networks

consist of two parts, ﬁrst part, work as typical CNN

building blocks, forward pass learning consist of two

layers linear ﬁltering and max pooling, with two terms

weight (W) is a Q × Q ﬁlter as shown in Fig. 7b,

and the scalar bias (B), with understanding back-

propagation, Also, the second layer is connected with

the CNN loss layer that contains the derivatives of

the loss (DZDY) of CNN output with respect to pos-

itive detection and negative detection in histopatho-

logical images. Here, the loss function can be repre-

sented as shown below: Where DZDY (Vedaldi and

Lenc, 2015) is the derivative of error or loss func-

tion, Pos is a positive blobs in histopathological im-

ages. Whereas, Neg is a negative blobs or background

structure as shown in Eq. 6, and Fig. 7a, Fig. 7b. Con-

volution neural networks associated with ground truth

information which there coordination give as the in-

formation of nuclei and background to learn CNN the

positive and negative location of nuclei detection as

shown in Fig. 6c. Where as, the second part looks at

learning two basic CNNs. The ﬁrst one is a simple

non-linear ﬁlter capturing particular image structures,

while the second one is a network that testing an im-

ages (using a variety of different colors).

Loss f unction =

Pos

∑

Pos

−

Neg

∑

Neg

(6)

The ﬁrst approach in this paper, which is obtain-

ing two data set inputs are used to train Tiny-CNNs

with different architectures, and gave the best results

accuracy of weight rate or bias from a new value,

where the weights are initialized randomly for using

(27 × 27 × 3) weight initialization. Also, the data

set which is used in this study, consist two types of

histopathological images which are UCSB-58 data

set and CRC-100 data set. We used of training in

convolutional neural network, to estimate the centers

of nuclei from the best model selection as shown in

Fig. 7d. In the UCSB-58 data set histopathological

images the network is trained with 58 images in data

set. Whereas, in the second data set CRC-100 net-

work trained for 100 images in data set. All the pre-

dictions of this algorithm have achieved a good result

to detect the center of nucleus location Fig. 7g. Our

study proposes to ﬁnd nuclei centers with Tiny-CNN.

To accomplish this task, we used the combined oper-

ation(combined the detection center of nuclei of FCM

semi automatic method with Tiny-CNN center detec-

tion). Our proposed study is detected the center of nu-

cleus by using conventional FCM clustering method,

on the original of histopathological image. This im-

age is used to compere with Tiny-CNN prediction

the center of nucleus. Also, apply a semi-automatic

FCM algorithm for extracted the perfect nucleus of

histopathological images, and ﬁnd the center of nu-

clei, to compere it also with Tiny-CNN result, Then,

we experement the result of combined this approach

with Tiny-CNN based automated method as shown in

Fig. 8. This scenario is applied on two data set CRC-

100 data set is accompanied with coordination (x,y)

ground truth information and UCSB-58 data set with

Automatic Nuclei Detection in Histopathological Images based on Convolutional Neural Networks

197

Figure 6: a) Detected positive nuclei centers for a sample

image in CRC-100 data set obtained by Tiny-CNN and b)

its negative image.

Figure 7: a) Original image from UCSB-58 data set, b) the

result of ﬁltering, c) Tiny-CNN output image detect of nu-

clei, d) ﬁnal output result of Tiny-CNN detect center of nu-

clei, e) FCM semiautomatic result, f) generated mask from

the FCM semiautomatic result combined with original im-

age g) used mask image with Tiny-CNN output the center

of the nuclei detection.

our contribution ground truth which is choose the co-

ordinates of (x,y) manually with the same manners of

CRC-100 data set ground truth.

4 RESULT AND DISCUSSION

In the detection of nuclei in histopathological im-

ages three approaches (Tiny-CNN, FCM, FCM-semi-

automatic) are evaluated and compared detection cen-

ter of nuclei in two different data set UCSB 58-set

and CRC 100-set. Our contributions include a dot op-

erator, by clicking on the center of a nuclei manu-

ally to produce ground truth information, for the ﬁrst

data set UCSB-58 data set to obtain the same manner

of ground-truth information in the CRC-100 data set.

This ground truth information can also be seen in this

study to detect the center of nuclei in histopathologi-

cal images. The ﬁrst approach uses detecting center of

cell by Tiny-CNN, second approach using FCM and

FCM semi-automatic algorithm to detect the center of

nuclei, then, that we compared these two approaches

to ﬁnd the best results for center of nucleus detection.

Afterwards, we combined center of nuclei result of

the Tiny-CNN and FCM semi-automatic approaches

to evaluate our model in detect center of nuclei.The

basic experimental results unit in this study, has used

the boundary of cell size to detected the correct nuclei

in related image. In the ﬁrst data set UCSB-58 data set

we stand a region information around 20 pixel as cir-

cular regions. Whereas, in the second data set CRC-

100 data set, we stand the circular regions within 10

pixels of each cell in the histopathological image. A

detection can be evaluated of cell center is identiﬁed

to be a true positive (TP) when it is matched within

the ground-truth information, otherwise, it is repre-

sent as a false positive (FP). Every true positive is

incorporated with the nearest ground-truth cell cen-

ter. The ground-truth cell centers that are not marched

by any detected results are represented as false nega-

tives (FN). The accuracy of nucleus detection is cal-

culated on a per nucleus basis on recording the true

positive (TP) (correctly identiﬁed nucleus success-

fully detected), false negative (FN) (incorrectly re-

jected nucleus), and false positive (FP) (identiﬁed non

nucleus objects found). In Fig. 9 the ground-truth re-

sults with Tiny-CNN results. In this study the auto-

mated method, contributors achieve results as shown

in Table 1. Focusing on the F-M measure we have also

deﬁned True Positive Rate (TPR) and Positive Predic-

tion Value (PPV). Based on the above deﬁnitions, we

observe the strongest performance reach to (FM be-

tween 0.769 and 0.979 ) as shown in Table 1. The

sensitivity or true positive rate (T PR) is estimated by

formula as in Eq. 7:

T PR =

T P

T P + FN

(7)

F1-score in Eq. 9 is represented as a metric of

overlap between the nuclei predicted and ground-truth

information. The F1-score is associated with true pos-

itive rate and precision prediction value in Eq. 8 .

PPV =

T P

T P + FP

(8)

F1 − score = 2 ×

T PR × PPV

T PR + PPV

(9)

DSC =

2T P

2T P + FP + FN

(10)

Dice Similarity Coefﬁcient (DSC) in Eq(10), is

used as a statistical validation measurement for eval-

uating the performance of both the predictability of

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

198

Figure 8: Proposed methodology, comparing FCM center

detection with Tiny-CNN center detection, combined FCM

and Tiny-CNN to detect center of nuclei.

Figure 9: a) An example image of UCSB-58 data set and, b)

red points represent outputs of nucleus center detection by

Tiny-CNN; green circles refer to ground truth information.

c) An example image of CRC-100 data set and, d) output of

detected nucleus centers Tiny-CNN and ground-truth.

manual segmentation and the overlap accuracy of au-

tomated segmentation of histopathological images.

All information about the experimental results

about detection of nuclei in histopathological images

in this study, are shown in Table 1. each different

method in this study give the different performance.

In this paper, the proposed methods has obtained of

true positive rate (TPR) of nucleus center detection

has reached to be between (0.863 - 0.964) at UCSB-

58 data set consisting of 58 images. In addition,

in this data set the (F-M or F1-score) has reached

to 0.979 in Tiny-CNN and 0.934 in the combined

FCM and Tiny-CNN and others FCM approaches are

reached between (0.89 - 0.933), whereas,in CRC data

set the true positive rate (TPR) has obtained results

between (0.547 - 0.621) and (F-M) has reached be-

tween (0.528 - 0.769), in this instance, the evaluating

of boundary in this study, has a directly effect the true

positive rate (TPR) as shown in Table 1.

5 CONCLUSIONS

In this study, we evaluated and compared fuzzy C-

means (FCM) clustering algorithm with tiny convo-

lutional neural network to detect center of nuclei in

histopathological images. To further explore nuclei in

data set images we combined and experimented two

approaches (Tiny-CNN and FCM semi-automatic) as

shown in Fig. 8. High results performance in (F-

M) reach to ( between 0.899 - 0.979) are observed

in the ﬁrst data University of California Santa Bar-

bara’s UCSB-58 set. However, we also observed in

the second, data set University of Warwick’s CRC-

100 set, the highest results performance in (F-M) be-

tween (0.741 - 0.769). In our models, provides ba-

sis for the qualitative form similarity between models

prediction and ground truth information to detect nu-

clei in histopathological images. The correct nuclei

is detected, according to ground-truth information as

shown in Fig. 9, it has shown that the prediction of

Tiny-CNN method with ground-truth information on

the UCSB-58 and CRC-100 data sets.

The future work of this study, we will con-

sider comparing performance of Tiny-CNN with

other CNN established method to detect nuclei in

histopathological images.

REFERENCES

Al-Kofahi, Y., Lassoued, W., Lee, W., and Roysam, B.

(2010). Improved automatic detection and segmen-

tation of cell nuclei in histopathology images. IEEE

Trans. on Biomedical Engineering, 57(4):841–852.

Albayrak, A. and Bilgin, G. (2016). Mitosis detection using

convolutional neural network based features. In IEEE

17th Int. Symposium on Computational Intelligence

and Informatics, CINTI’16, pages 335–340. IEEE.

Albayrak, A. and Bilgin, G. (2018). Automatic cell seg-

mentation in histopathological images via two-staged

superpixel-based algorithms. Medical & Biological

Engineering & Computing, pages 1–13.

Braz, E. F. and Lotufo, R. (2017). Nuclei detection us-

ing deep learning. Simp

osio Brasileiro de e Proces-

samento de Sinais, pages 1059–1063.

Gelasca, E. D., Obara, B., Fedorov, D., Kvilekval, K., and

Manjunath, B. (2009). A biosegmentation benchmark

for evaluation of bioimage analysis methods. BMC

Bioinformatics, 10(1):368.

Glenn, T. C., Zare, A., and Gader, P. D. (2015). Bayesian

fuzzy clustering. IEEE Transactions on Fuzzy Sys-

tems, 23(5):1545–1561.

Haﬁane, A., Bunyak, F., and Palaniappan, K. (2008). Fuzzy

clustering and active contours for histopathology im-

age segmentation and nuclei detection. In Int. Conf.

on Advanced Concepts for Intelligent Vision Systems,

pages 903–914. Springer.

Automatic Nuclei Detection in Histopathological Images based on Convolutional Neural Networks

199

Table 1: Nuclei detection results with UCSB-58 and CRC-100 histopathological image data sets by using Tiny-CNN and

FCM semiautomatic, automatic method, and, combined FCM semiautomatic with Tiny-CNN.

Data Set TPR PPV F-M DSC

UCSB-58 set with Tiny-CNN 0.964 0.996 0.979 0.979

UCSB-58 set with FCM semiautomatic 0.881 0.994 0.933 0.933

UCSB-58 set with FCM automatic 0.828 0.994 0.899 0.899

CRC-100 set with Tiny-CNN 0.621 0.989 0.769 0.769

CRC-100 set with FCM semiautomatic 0.466 0.994 0.617 0.617

CRC-100 set with FCM automatic 0.451 0.994 0.528 0.528

UCSB-58 set with combined FCM and Tiny-CNN 0.863 0.994 0.934 0.934

CRC-100 set with combined FCM and Tiny-CNN 0.547 0.967 0.741 0.741

Hatipoglu, N. and Bilgin, G. (2017). Cell segmentation

in histopathological images with deep learning algo-

rithms by utilizing spatial relationships. Medical &

Biological Engineering & Computing, 55(10):1829–

1848.

John, J., Nair, M. S., Kumar, P. A., and Wilscy, M. (2016).

A novel approach for detection and delineation of cell

nuclei using feature similarity index measure. Biocy-

bernetics and Biomedical Engineering, 36(1):76–88.

Kainz, P., Urschler, M., Schulter, S., Wohlhart, P., and Lep-

etit, V. (2015). You should use regression to detect

cells. In International Conference on Medical Im-

age Computing and Computer-Assisted Intervention,

pages 276–283. Springer.

Khoshdeli, M., Cong, R., and Parvin, B. (2017). Detection

of nuclei in H&E stained sections using convolutional

neural networks. In IEEE EMBS Int. Conf. on Biomed-

ical & Health Informatics, BHI’17, pages 105–108.

Kong, H., Akakin, H. C., and Sarma, S. E. (2013). A gen-

eralized laplacian of gaussian ﬁlter for blob detection

and its applications. IEEE Transactions on Cybernet-

ics, 43(6):1719–1733.

LeCun, Y., Kavukcuoglu, K., Farabet, C., et al. (2010).

Convolutional networks and applications in vision. In

ISCAS, volume 2010, pages 253–256.

Moita, A. S., Jacinto, F., and Moreira, A. L. N. (2018).

Design, test and fabrication of a droplet based mi-

croﬂuidic device for clinical diagnostics. In Proceed-

ings of the 11th Int. Joint Conference on Biomed-

ical Engineering Systems and Technologies-Volume

1: BIODEVICES, BIOSTEC’18, pages 88–95. IN-

STICC, SciTePress.

Phung, S. L. and Bouzerdoum, A. (2009). Matlab library

for convolutional neural networks. University of Wol-

longong, Tech. Rep., URL: http://www. elec. uow. edu.

au/staff/sphung.

Sirinukunwattana, K., Raza, S. E. A., Tsang, Y.-W., Snead,

D. R., Cree, I. A., and Rajpoot, N. M. (2016). Local-

ity sensitive deep learning for detection and classiﬁ-

cation of nuclei in routine colon cancer histology im-

ages. IEEE Trans. on Medical Imaging, 35(5):1196–

1206.

Sornapudi, S., Stanley, R. J., Stoecker, W. V., Almubarak,

H., Long, R., Antani, S., Thoma, G., Zuna, R., and

Frazier, S. R. (2018). Deep learning nuclei detection

in digitized histology images by superpixels. Journal

of Pathology Informatics, 9.

Suganya, R. and Shanthi, R. (2012). Fuzzy c-means

algorithm-a review. Int. Journal of Scientiﬁc and Re-

search Publications, 2(11):1.

Vedaldi, A. and Lenc, K. (2015). Matconvnet: Convolu-

tional neural networks for matlab. In Proceedings of

the 23rd ACM International Conference on Multime-

dia, pages 689–692.

Vink, J. P., Van Leeuwen, M., Van Deurzen, C., and

De Haan, G. (2013). Efﬁcient nucleus detector

in histopathology images. Journal of Microscopy,

249(2):124–135.

Xu, H., Lu, C., Berendt, R., Jha, N., and Mandal, M.

(2017). Automatic nuclei detection based on gener-

alized laplacian of gaussian ﬁlters. IEEE Journal of

Biomedical and Health Informatics, 21(3):826–837.

Xu, J., Xiang, L., Liu, Q., Gilmore, H., Wu, J., Tang, J.,

and Madabhushi, A. (2016). Stacked sparse autoen-

coder (SSAE) for nuclei detection on breast cancer

histopathology images. IEEE Transactions on Med-

ical Imaging, 35(1):119–130.

BIOSIGNALS 2019 - 12th International Conference on Bio-inspired Systems and Signal Processing

200