Comparison of HOG, MSER, SIFT, FAST, LBP and CANNY Features For Cell Detection in Histopathological Images

Download as pdf or txt
Download as pdf or txt
You are on page 1of 6

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/324983531

Comparison of HOG, MSER, SIFT, FAST, LBP and CANNY features for cell
detection in histopathological images

Article · May 2018


DOI: 10.29042/2018-3321-3325

CITATIONS READS

4 1,836

2 authors:

Şaban Öztürk Bayram Akdemir


Amasya University Selcuk University
40 PUBLICATIONS   118 CITATIONS    48 PUBLICATIONS   278 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Artificial Intelligence in Covid-19 View project

related to code security View project

All content following this page was uploaded by Şaban Öztürk on 07 May 2018.

The user has requested enhancement of the downloaded file.


DOI 10.29042/2018-3321-3325
Helix Vol. 8(3):3321-3325

Comparison of HOG, MSER, SIFT, FAST, LBP and CANNY features


for cell detection in histopathological images
1
Şaban Öztürk, 2 Bayram Akdemir
1, 2
Electrical and Electronics Engineering Department, Selçuk University, Konya, Turkey
Email: [email protected], [email protected]

Received: 10th Feb 2018, Accepted: 18th March 2018, Published: 31st April 2018

Abstract segmentation [7], mean-shift clustering based method


Cell segmentation and counting has a very important [8].
role in diagnosing diseases and in the treatment In the image segmentation or classification
process. But the complexity of the histopathological process using image processing techniques, features
images and the differences in cell groups make this obtained from images are affecting success. For this
process very difficult, even for an expert. In order to reason different features are used for different
facilitate this process, analysis of histopathological problems. The characteristics obtained by different
images is performed by using computer vision feature extraction methods applied on the same
methods. This paper presents the use of different problem can lead to different success rates. For this
feature extraction methods for cell detection in reason, many feature extraction methods have been
histopathological images and the comparison of the proposed in the literature. For example, feature
results of these algorithms. For this reason, HOG, extraction methods such as SIFT and SURF are used
MSER, SIFT, FAST, LBP and CANNY feature for interest region detection, and LBP and LPQ are
extraction algorithms are used. The aim of the study preferred for texture classification operations [9].
is to determine cells using different feature extraction In this study, the success of well-known feature
methods and to determine which of these feature extraction methods for cell detection in
extraction algorithms will be more successful. Firstly, histopathological images is investigated. For this
image pre-processing has been applied to clear the purpose, HOG, MSER, SIFT, FAST, LBP feature
noises in the histopathological images. Then, feature extraction methods and canny edge detection
extraction algorithms are applied to image, algorithm are used. In order to detect the cells through
respectively. Finally, the successes of different the features obtained from the histopathological image
feature extraction algorithms have been compared. using the mentioned algorithms, the feature sets need
only cover the cells. For this reason, when applied to
Keywords: HOG, MSER, SIFT, FAST, LBP, canny, raw images, they fail because of disturbing factors
histopathological image, cell counting. such as noise. In this study, firstly RGB color space to
Introduction HSV color space conversion is done. Then, noise in
Computer aided diagnosis studies in biomedical the V component of the HSV color space is cleaned
engineering field are very useful for human health and using wiener filter [10]. Then, image sharpening is
diagnosis of diseases. In particular, the state of cellular applied to the image. With the help of pre-processing,
structures and cell forms present useful information feature extraction algorithms have increased the
about many diseases and cancer types [1]. For this chances of catching only the cells. In the next step,
reason, cell forms and behaviors in histopathological HOG [11], MSER [12], SIFT [13], FAST [14], LBP
images have become one of the main research areas. [15] and canny [16] algorithms are applied
However, because of the negative and compelling respectively. Finally, cells are identified with the help
properties of these images, such as unwanted light of extracted features and the success of each algorithm
effects, low contrast, differences in cell shapes etc. is compared with other algorithms.
analysis takes a long time. Again for this reason, even
if these images are evaluated by an expert, System Overview:
histopathological image analysis is quite challenging This study involves pre-processing step and well-
known feature extraction algorithms. Well-known
and time-consuming. Also, it requires skill and
feature extraction algorithms are applied to
experience [2]. In order to overcome all these histopathological images and the success rates of these
difficulties, a computer-assisted automatic cell algorithms are determined. In the proposed system,
detection system is needed. Image processing the image is first prepared for feature extraction. For
algorithms which are used heavily in these systems this, the histopathological image is transformed into
and directly affect the success of the system have an HSV color space and inserted into various
important effect in this context [3, 4]. Some studies in preprocessing stages. Then feature extraction
the literature for cell detection from histopathological algorithms are applied to the image in sequence. The
images are; compact Hough and radial map [5], mentioned processes are shown in Fig 1.
logistic regression classifier [6], watershed

3321 Copyright © 2018 Helix ISSN 2319 – 5592 (Online)


Helix Vol. 8(3):3321-3325

The Wiener filter [18] is used to remove noise from


the V component. The Wiener filter obtains the
Pre-Processing Feature Extraction statistical properties of the image according to the
determined neighborhood value. For a great
Histopathological HOG neighborhood value, the smoothing becomes less and
İmage
the smoothing is big for a small neighborhood value.
Finally, the image is sharpened to increase the clarity
MSERS of the cells. Image after pre-processing is shown in
Conversion from
RGB to HSV Fig. 3.
SIFT Fig. 3. V component after pre-processing
Obtaining V value
from HSV
FAST

Wiener filter
LBP

Image sharpening Canny

Fig. 1. Proposed system overview

Pre-Processing of Histopathological images


The main algorithms in image processing applications
are mostly sensitive to noise and light variations. For
this reason, processing raw images prevents the
achievement of desired success. Image pre-processing Feature Extrction
techniques are used to avoid noise and illumination For object recognition and classification operations,
fluctuations in images and they are used to reduce the extracting properties from the image and using these
effect on image processing algorithms [17]. The properties is the basic procedure. Basically, feature
noises in the images are usually of different types. For extraction is done according to orientation, scale and
this reason, problem-specific or image-specific image transformations [19]. Many algorithms are used with
preprocessing techniques are used. gray level images, and some algorithms can be used
for color images. Extraction of efficient features is
In this study, an image preprocessing solution is very important for image processing problems.
presented for low contrast, color problems caused by However, efficient features may be different for each
H & E stained and various noises in histopathological image problem. For this reason many feature
images. First, the image is converted from RGB color extraction methods are presented. In this study, the
space to HSV color space. Then, V component of success of well-known feature extraction methods is
HSV color space is selected for further processing. V tested on histopathologic images.
component is shown in Fig 2.
A. HOG Feature Extraction
Fig. 2. V component of HSV color space
Histogram of gradients (HOG) specifies the
magnitude and orientation parameters of feature
regions in an image. These parameters are used to
determine the objects in the image and image
matching. They are independent of the size.
Magnitude m(x,y) and orientation θ(x,y) are calculated
as in Eq. 1. and Eq. 2.

m  x, y   f x  x, y   f y  x, y 
2 2

(1)
f y  x, y 
  x, y   arctan
f x  x, y 
(2)
A feature image which is trained by cells for cell
detection in histopathological images is shown in
Figure 4. This feature matrix is applied to the test
images to identify the cells.

3322 Copyright © 2018 Helix ISSN 2319 – 5592 (Online)


Helix Vol. 8(3):3321-3325

Fig. 4. HOG feature for cell detection E. LBP Feature Extraction


Local binary pattern (LBP) labels each pixel by
thresholding it according to nxn neighborhood. A
feature histogram representing the visual features is
created with the obtained features. This feature
histogram is searched in other images and similarities
are determined. When similarity is determined,
feature histograms are determined in other images and
the similarities between two feature histograms are
calculated.
F. Canny Edge Detector
Canny edge detection algorithm is one of the most
commonly used edge detection methods. The work of
this algorithm is summarized as follows: firstly, the
noise in the image is cleared. If the noises are not
cleaned, the correct edges cannot be obtained. Then, a
B. MSER Feature Extraction two-dimensional derivative of the image intensity in
the x and y directions (Gx, Gy) are calculated. Finally,
Maximally stable extremal region (MSER) the gradient of the view is calculated by adding Gx
extracts a large number of affine covariant regions and Gy values [21].
from the image. These features are usually elliptical
regions. In the MSER algorithm, the threshold values Results and Discussion
are constantly changed to evaluate the areas whose In this study, the performance of the feature extraction
shape is preserved. Using this algorithm, the capture algorithms in the cell detection process is examined.
of elliptical features in the image produces very First of all, random image patches are cut from
successful results. For this reason, this algorithm has histopathological images. These image patches are
been chosen to be used for cell detection. cleaned with the preprocessing algorithm and are
C. SIFT Feature Extraction made feasible for feature extraction process. In the
experiments, 12 pieces of histopathological image are
The Scale invariant feature transform (SIFT)
examined. The cells of the image patches are marked
algorithm is used to capture local features in the
image. The state of the specified properties is not by an expert. The total number of cells in the marked
affected by scaling, rotation, translation. First, local image patches is determined. Then, feature extraction
maxima and minima are determined by Gaussian process is performed according to their own capability
function with different values. Equation 3 is used for and characteristics. The purpose of this process is to
this operation. determine only the regions containing the cells. The
performance of the well-known feature extraction
D( x, y,  )  G ( x, y, t )  G ( x, y,  )  I ( x, y) algorithms are shown in Table 1.
(3)
where t represents constant factor, σ represents TABLE 1. COMPARISON OF WELL-KNOWN FEATURE
Gaussian difference value. Then, each pixel is EXTRACTION ALGORITHMS
compared with the neighboring pixels and the regions Undetec Total
with low contrast values are eliminated by examining Correct Uncorrected
ted Cells Cell
the key points. The original paper written by Lowe Cells Cells
number
[20] can be examined for detailed information.
HOG 2748 76 1176 3924
D. FAST Feature Extraction
Features from accelerated segment test (FAST) MSER 3540 22 384 3924
determines the interest points in the image. Interest SIFT 2903 428 1021 3924
points carry local information about the content of
interest and can be repeated for different images. FAST 3064 294 860 3924
Thanks to these features, they can be used for object LBP 3489 158 435 3924
detection, object tracking and scene matching
operations. This algorithm checks whether each pixel CANNY 3322 38 602 3924
is an interest point. The selected pixel intensity value
is evaluated according to the intensity value of the In table 1, correctly identified cells are called "correct
neighboring 16 pixels. For a pixel can be an intensity cell", cells that cannot be detected correctly are called
point, the value of this pixel must be different from the "undetected cells" and regions that are marked as cells
neighboring pixels around it after the threshold when they are not cells are called "uncorrected cells".
operation.

3323 Copyright © 2018 Helix ISSN 2319 – 5592 (Online)


Helix Vol. 8(3):3321-3325

Original Ground Truth HOG MSER SIFT FAST LBP CANNY

Fig: 5 HOG feature for cell detection

The features determined by well-known feature extraction Conclusion


algorithms are shown in Figure 5. The MSER algorithm In this study, the success of well-known feature extraction
produces better results than other algorithms. It could detect methods is investigated for cell detection in histopathological
many cells and mark around. But the cells that are close to each images. It has been understood that feature extraction algorithms
other are perceived as a single cell by MSER algorithm. The cannot achieve the desired success on raw images. For this
LBP algorithm is another algorithm that can accurately reason, preprocessing is applied before image properties are
determine the locations of cells. But too many false positives are extracted. Then feature extraction algorithms are applied to the
marked as positive. The Canny edge detection algorithm cleaned image. Each feature extraction algorithm operates on its
determines the edges of the cells very successfully. However, own characteristic feature. In order for the comparison process
since many regions have an intensity difference, non-cell
to be objective, these characteristics are preserved and applied
regions are drawn as a cell by canny algorithm. The HOG
algorithm is not very successful at catching all the cells, but the in the same conditions. At the end of the comparison process,
false markup rate is low compared to other algorithms. For this MSER, LBP and canny algorithms detect cells with high
reason, it is more useful than LBP and canny algorithms. The success. However, the LBP and canny algorithms incorrectly
FAST algorithm puts more than one sign on the same cell. The mark many regions. In this case, the accuracy rate is low. The
SIFT algorithm cannot detect many cells and marks many non- HOG algorithm is more advantageous than the LBP and canny
cell regions. algorithms because it has a low error rate when it has high

3324 Copyright © 2018 Helix ISSN 2319 – 5592 (Online)


Helix Vol. 8(3):3321-3325

success in determining the cells. In line with this information, (ICPR), 2012 21st International Conference on (pp. 1363-
the MSER algorithm can be used as the first option and the HOG 1366). IEEE.
algorithm can be used as the second option. But, the selection [10] Benesty, J., Chen, J., Huang, Y. A., & Doclo, S. (2005).
Study of the Wiener filter for noise reduction. In Speech
process may vary depending on the application. Enhancement (pp. 9-41). Springer Berlin Heidelberg.
[11] Dalal, N., & Triggs, B. (2005, June). Histograms of
oriented gradients for human detection. In Computer
References Vision and Pattern Recognition, 2005. CVPR 2005. IEEE
[1] Mao-Jun, S., Zhao-bin, W., Hong-Juan, Z., & Yi-de, M. Computer Society Conference on (Vol. 1, pp. 886-893).
(2008, March). A new method for blood cell image IEEE.
segmentation and counting based on PCNN and autowave. [12] Turki, H., Halima, M. B., & Alimi, A. M. (2017, March).
In Communications, Control and Signal Processing, 2008. A Hybrid Method of Natural Scene Text Detection Using
ISCCSP 2008. 3rd International Symposium on (pp. 6-9). MSERs Masks in HSV Space Color. In Ninth International
IEEE. Conference on Machine Vision (pp. 1034111-1034111).
[2] Mölder, A., Sebesta, M., Gustafsson, M., Gisselson, L., International Society for Optics and Photonics.
Wingren, A. G., & Alm, K. (2008). Non‐invasive, label‐ [13] Ma, B., Ban, X., Wang, Y., & Liu, X. (2016, August).
free cell counting and quantitative analysis of adherent Recognition of blacklisted vehicle based on SIFT feature.
cells using digital holography. Journal of microscopy, In Cloud Computing and Intelligence Systems (CCIS),
232(2), 240-247. 2016 4th International Conference on (pp. 456-460). IEEE.
[3] Bhagavathi, S. L., & Thomas Niba, S. (2016). An
Automatic System for Detecting and Counting RBC and [14] Wang, D., Hou, X., Xu, J., Yue, S., & Liu, C. L. (2017).
WBC using Fuzzy Logic. ARPN Journal of Engineering Traffic sign detection using a cascade method with fast
and Applied Sciences, 11(11), 6891-94. feature extraction and saliency test. IEEE Transactions on
Intelligent Transportation Systems.
[4] Sommer, C., Straehle, C., Koethe, U., & Hamprecht, F. A.
(2011, March). Ilastik: Interactive learning and [15] Stella, X. A., Sivagangai, T., & Sujatha, I. D. N. (2016).
segmentation toolkit. In Biomedical Imaging: From Nano Performance Analysis of GFE, HOG and LBP Feature
to Macro, 2011 IEEE International Symposium on (pp. Extraction Techniques using kNN Classifier for Oral
230-233). IEEE. Cancer Detection. Journal of Network Communications
and Emerging Technologies (JNCET) www. jncet. org,
[5] Barber, P. R., Locke, R. J., Pierce, G. P., Rothkamm, K., & 6(7).
Vojnovic, B. (2007, January). Gamma-H2AX foci [16] Goulart, J. T., Bassani, R. A., & Bassani, J. W. M. (2017).
counting: image processing and control software for high- Application based on the Canny edge detection algorithm
content screening. In Proceedings of SPIE (Vol. 6441, pp. for recording contractions of isolated cardiac myocytes.
64411M-1). Computers in biology and medicine, 81, 106-110.
[6] Liimatainen, K., Ruusuvuori, P., Latonen, L., & Huttunen,
H. (2016, April). Supervised method for cell counting from [17] Heseltine, T., Pears, N., & Austin, J. (2002, July).
bright field focus stacks. In Biomedical Imaging (ISBI), Evaluation of image pre-processing techniques for
2016 IEEE 13th International Symposium on (pp. 391- eigenface based face recognition. In Proc. of the Second
394). IEEE. International Conference on Image and Graphics, SPIE
(Vol. 4875, pp. 677-685).
[7] Veta, M., van Diest, P. J., Kornegoor, R., Huisman, A.,
Viergever, M. A., & Pluim, J. P. (2013). Automatic nuclei [18] Pratt, W. K. “Generalized Wiener filtering computation
segmentation in H&E stained breast cancer histopathology techniques.” IEEE Transactions on Computers, 100(7),
images. PloS one, 8(7), e70221. 1972, 636-641.
[8] Qi, X., Xing, F., Foran, D. J., & Yang, L. (2012). Robust
[19] Van De Weijer, J., & Schmid, C. (2006). Coloring local
segmentation of overlapping cells in histopathology feature extraction. Computer Vision–ECCV 2006, 334-
specimens using parallel seed detection and repulsive level 348.
set. IEEE Transactions on Biomedical Engineering, 59(3), [20] D. Lowe, “Distinctive Image Features from Scaleinvariant
754-765. VKey Points”, IJCV, 60(2), 91-110, 2004
[9] Kannala, J., & Rahtu, E. (2012, November). Bsif: [21] Öztürk, Ş., & Akdemir, B. (2015). Comparison of edge
Binarized statistical image features. In Pattern Recognition detection algorithms for texture analysis on glass
production. Procedia-Social and Behavioral Sciences, 195,
2675-2682.

3325 Copyright © 2018 Helix ISSN 2319 – 5592 (Online)

View publication stats

You might also like