A Comparative Study of BRISK, ORB and DAISY Features for Breast

Cancer Classiﬁcation

Ghada Ouddai, Ines Hamdi and Henda Ben Ghezala

RIADI Laboratory, National School of Computer Science (ENSI), University of La Manouba, La Manouba, Tunisia

Keywords:

Histopathological Image Processing, Feature Extraction, Binary Robust Invariant Scalable (BRISK), Oriented

FAST and Rotated BRIEF (ORB), DAISY Descriptor, Bag-of-Features (BoF), Machine Learning.

Abstract:

Medical data analysis is one of the most emergent ﬁelds over the past decades. In Digital histopathology,

images are analysed, mainly, to detect disease or tumors and identify their types and grade. One of the most

used practices in this ﬁeld is the feature extraction. In this paper, we propose the application of BRISK, ORB

and BRISK/DAISY on RGB histological images. The purpose of this work is to recognise the breast tumor

type (benign or malignant). These features extractors are combined with BoF by kmeans and SVM. A limited

amount of images is used during the training of the system. Out of the three methods, Color-BRISK/BoF/SVM

solution gave the best accuracy value (72.5%) while Color-ORB/BoF/SVM was the fastest.

1 INTRODUCTION

Histopathology, which is also known as pathological

histology, is a bio-medical ﬁeld that offers a useful

techniques for cancer and disease detection and grad-

ing. Histology and histopathology share the same

sample preparation, called histological process, and

the same study tool: the microscope. The difference

between these two sub-ﬁelds is the purpose of study:

in histology, samples as analyzed to observe the cells

morphological development, on the other side, a sam-

ple study in histopathology is performed to detect

abnormal tissues and diseases. When examined for

pathological purposes, a histological sample can be

very effective for detection of tumors as well as deﬁn-

ing its nature and grade.

Following the digitalization of medical data, an

emergence of AI tools applications to these latter has

been observed. In digital histopathology, the main

domain data is the histopathological image, which is

generated by scanning a given specimen. Depend-

ing on the tools, stains and staining techniques used

during the histological process, the image processing

method is selected. In fact, in histology, their are vari-

ous tools for sample cutting, preparation and staining.

The three known-to-date staining techniques are: his-

tochemistry (HC), immunohistochemistry (IHC) and

immunoﬂuorescence (IF), for each one there are hun-

dreds of possible stains. The choice of stains depends

on the target cell and study context; each stain or

stains combination allows the emphasis and highlight

of certain morphological parts, the frequently used

ones are Eosin (E), Hematoxylin (H) and their com-

bination (H&E).

Over the past years, there was a huge number

of researches and attempts to create the perfect au-

tonomous Computer-Assisted Diagnosis (CAD) sys-

tem for disease and cancer detection and grading us-

ing histopathological images. In parallel to that, nu-

merous works focused on the images retrieval and/or

registration. Our study of the state-of-the-art works

in histopathological image analysis ﬁeld, such as pre-

sented in the papers (Azevedo Tosta et al., 2017),

(Das et al., 2020), (Gurcan et al., 2009), (Irshad et al.,

2014), (Komura and Ishikawa, 2018), (Li et al., 2020),

(Ai et al., 2021), showed that the majority of CADs

proposed in this ﬁeld depends on deep learning meth-

ods. These latter offer great classiﬁcation results how-

ever there are some limitations to them:

• First, to achieve good result, a considerable num-

ber of labelled data should be used. The more

images is analysed, the good performance is ob-

tained.

• Second, deep learning methods need powerful

computation machines. GPU-based calculation

offers fast CNN and RNN training however, if the

experiments data-set is very large, a memory over-

ﬂow can occur. In other hand, CPU-based calcu-

lation is slow but the memory is unlimited.

964

Ouddai, G., Hamdi, I. and Ben Ghezala, H.

A Comparative Study of BRISK, ORB and DAISY Features for Breast Cancer Classiﬁcation.

DOI: 10.5220/0011902200003411

In Proceedings of the 12th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2023), pages 964-970

ISBN: 978-989-758-626-2; ISSN: 2184-4313

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

The available labelled histopathological data-sets

are mainly consisted of HC H&E slides. There are

some IHC data-sets but, to the best of our knowl-

edge, there is no dedicated IF images collections.

The majority of these image-bases regroup breast tu-

mor slides either by type (benign, malignant) or by

grade. In this work, to classify breast tumoral cells,

we use and compare three features extractors: Bi-

nary Robust Invariant Scalable (BRISK) (Leuteneg-

ger et al., 2011), Oriented FAST and Rotated BRIEF

(ORB) (El-Hallak and Lovell, 2013) and BRISK-

keypoints/DAISY-descriptors. These methods are

applied to RGB images rather than gray-scale im-

ages; this allows the exploitation of color informa-

tion. For each feature extractor, an encoding on Bag-

of-Features (BoF) by kmeans/frequency histogram is

performed before the last step of classiﬁcation by

SVM (Support Vector Machine). A CPU-based cal-

culation and a limited data-set are used in the experi-

ments. The main purpose of this work is to ﬁnd which

feature extractor is the fastest and the more accurate

in identifying the tumor type.

This paper is organized as follows: in section

2, we give a quick overview of literature works and

features extraction applications in digital histopathol-

ogy, then, in section 3, we introduce and explain our

approach in details beginning by the pre-processing

method used till the classiﬁcation system. The pro-

posed approach is evaluated in section 4 where all

the speciﬁcations of computation architecture, data-

set and test results are listed. In section 5, we give

a conclusion of this paper and a perspective of future

works.

2 BACKGROUND

2.1 Low-Level Information Detection

In digital histopathology, image processing tech-

niques are used mainly for segmentation or Regions-

of-Interest (ROI) detection. Features extraction is

less used. In (

Ozt

urk and Akdemir, 2018), the au-

thors study the efﬁciency of the combination of dif-

ferent feature extractors with a variety of classi-

ﬁer. For the texture characteristics, Gray-Level Co-

occurrence Matrix (GLCM), Gray-Level Run Length

Matrix (GLRLM) and Segmentation-based Fractal

Texture Analysis (SFTA) are calculated. For the lumi-

nance features, the authors used Local Binary Pattern

(LBP). The Local Binary Gray Level Co-occurrence

Matrix (LBGLCM) is also calculated for common

texture/luminance features. The evaluation of these

methods was performed by the authors using some

common classiﬁers such as Support Vector Machine

(SVM), K-Nearest Neighbors (KNN), Linear Dis-

criminant Analysis (LDA) and Boosted Trees. The

best performance was achieved by the SFTA/Boosted

Trees system.

Local Binary Pattern (LBP) was used in (Ku-

mar et al., 2018) combined with Bag-of-Visual-Words

(BoVW) and SVM; a comparison with LBP deep

features was established. A variant of LBP, named

mrcLBP which consists of calculating LBP on each

RGB channel separately, was used in (Simon et al.,

2018). Other than that, KAZE features were used in

(Sanchez-Morillo et al., 2018) to classify breast can-

cer H&E-stained images. In (Popovici et al., 2016),

the authors propose to use directly the clustering by

kmeans to construct a local Bag-of-Features (BoF)

from the image, named code blocks. these latter are

jointed to tumor size, grade and gene expression.

Scale-Invariant Feature Transform (SIFT) (Lowe,

1999) was used in (Li et al., 2019b) and (Li et al.,

2019a) alongside other methods to prepare cervical

histopathological images for classiﬁcation. In (Irshad

et al., 2013), to detect mitosis from H&E stained im-

ages, the authors proposed an system based on SIFT

and texture features to detect the key-points from R

and B channels. In (Bukała et al., 2020), the authors

proposed the exploitation of color information by us-

ing and comparing various Color-SIFT. Similar pro-

cedure was performed in (Ouddai et al., 2023) where

the authors used RGB-SIFT to classify breast cancer

using a small data-set.

2.2 Databases

For CAD, databases are needed to train the ma-

chine/deep learning systems. Digital histopatholog-

ical databases regroup similar images: the study con-

text, size (Whole slide image WSI, patches or reg-

ular sized images) and stains used (histochemistry

(HC), immunohistochemistry (IHC), immunoﬂuores-

cence (IF), Eosin (E), Hematoxylin (H) or their com-

bination (H&E)) must be the same. Databases in gen-

eral need to be labeled by ﬁeld experts. In the case of

digital histopathology, the size of image, stains used

and studied cells must be provided with the images,

some additional details such as: age, gender, health

situation. . . etc. can be useful in some studies. In Tab.

1, we present some of the existing histopathological

databases.

A Comparative Study of BRISK, ORB and DAISY Features for Breast Cancer Classiﬁcation

965

Table 1: Examples of Histopathological Image Databases.

Name Type Stains used Cell Dataset size

BreCaHAD (Aksac et al., 2019) Regular images H&E Breast 162

(Wang et al., 2022) WSI H&E ovarian Cancer 288

BreakHis (Spanhol et al., 2016) Regular image H&E Breast 7909 images

Medisp HICL (Glotsos et al., 2008), Regular image H&E/IHC Brain, Breast 3870 images

(Kostopoulos et al., ), Larynx

(Ninos et al., 2016)

Camelyon17 (Litjens et al., 2018) WSI H&E Breast 1339 images

22591

KIMIA Path24 (Shaﬁei et al., 2021) WSI patches H&E/IHC / (train patches)

1325 (test patches)

3 METHODOLOGY

As stated before, in digital histopathology, color in-

formation is very important; it indicates the histolog-

ical process staining results. The choice of staining

techniques and stains depends on the cells or disease

to detect and study, in fact, for each case exists one or

multiple adequate pigmentation method. The choice

of this latter is an important step in the histological

process.

In this paper, we propose the modiﬁcation of reg-

ular features descriptors (in our case: BRISK, ORB

and DAISY) to extract key-points from RGB chan-

nel rather than gray-scale single channeled image.

Alongside that, we use an encoding method to re-

group descriptor into clusters of same nature. In the

ﬁnal step, a supervised classiﬁcation method is used

to determinate and interpret the input image nature.

The context of our study is the classiﬁcation of breast

tumor by type (benign/malignant). The main purpose

of this work is to present answers to the following re-

search questions:

• Between color-BRISK descriptor, color-ORB

descriptor and color-BRISK-keypoints/DAISY-

descriptor, which method is the more appropriate

to histopathological images?

• When dealing with a limited database, which

method can offer a better slide classiﬁcation

scores?

• For each method and for a CPU-based calcula-

tion, what is the maximal execution time to be ex-

pected?

3.1 Pre-Processing

In any computer vision sub-domain, the preparation

of input image for further processing, analysis and

interpretation is really important; the pre-processing

tools and method must be chosen thoroughly. In digi-

tal histopathology in particular, the nature of image is

delicate due to the morphological textures of cells and

tissues. The most frequent noise that can occur dur-

ing the scanning of histological slides are green hues

or shadows and luminance unbalance.

The database selected in our study consists mainly

of HC H&E stained slides. When analysing these

samples, the ﬁrst global remark is that the majority

of images contain a green shadow; its intensity dif-

fers from image to another. To remedy to this prob-

lem and eliminating the hue without loosing impor-

tant morphological textures, we use a lightweight pre-

processing method; we chose the bias and gain func-

tion (see Eq. 1) to correct the luminance and contrast.

Out put(i, j) = α ∗ Input(i, j) + β (1)

In the deﬁnition of the equation above, the param-

eter α and β are ﬁxed in an experimental way: α con-

trols the contrast while β controls the brightness. For

the parameter α, its value must be between 1 and 3;

if α < 0, the result image colors will be compressed.

For the parameter β, its value should range between 0

and 100. The procedure of selection of these latter’s

values depends on the nature of the studied images. In

the case of HC H&E stained slides, we noticed that a

green shadow appears more often than in IHC images.

Our ﬁrst tentative of adjusting the images quality us-

ing the same parameter’s values on the three chan-

nels was unsatisfactory. This latter led to the fading

of some important morphological and color details.

After numerous experiments, we found that ad-

justing each channel of the RGB image separately is

the best solution to obtain well calibrated contrast and

brightness. For the HC H&E slides, we ﬁx α = 1.2

and β = 25 for the R and B channels, for the G chan-

nel, we ﬁx α = 1.2 and β = −25. In the case of IHC

image, we ﬁx α = 1.1 and β = 10, for the three chan-

nels. The pre-processing results are shown in Fig. 1.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

966

Figure 1: Results of our pre-processing methods – Exam-

ples of H&E and IHC images of Medisp HICL database

((Glotsos et al., 2008), (Kostopoulos et al., ) and (Ninos

et al., 2016)).

3.2 Features Extraction

The main purpose of this work, as stated before, is to

apply, evaluate and compare three methods for fea-

tures extraction. These latter are slightly modiﬁed

to operate on RGB image instead of gray-scale im-

ages. With this, we ensure the use of the histolog-

ical staining information as well as the morphologi-

cal textures. In our work, we chose the method: Bi-

nary Robust Invariant Scalable (BRISK) (Leuteneg-

ger et al., 2011), Oriented FAST and Rotated BRIEF

(ORB) and a combination of BRISK key-points and

DAISY descriptors.

3.2.1 Features Extraction by Color-BRISK

The original Binary Robust Invariant Scalable

(BRISK), as presented in (Leutenegger et al., 2011)

is applied to the gray-scale image. This method al-

lows the detection of local key-points and their de-

scriptors construction. The results are rotation, scal-

ing and translation invariant. BRISK follows the same

strategy as SIFT (Lowe, 1999) while being faster.

Gray-scale image is the result generated from a di-

rect transformation of the original image and combi-

nation of its three RGB channels. Some loss of color

information can occur following this conversion. To

make the most use of such details, we propose a mod-

iﬁcation to the original BRISK:

• For each RGB channel, apply BRISK and extract

the key-points.

• Generate the descriptor vector by BRISK for each

channel.

• Concatenate the three descriptor vectors into one

vector.

By this methodology, each channel of the RGB

image is considerate by BRISK as a gray-scale im-

age. In the end, and by combining the three descrip-

tors, we are sure to conserve the color information

of each channel. The ﬁnal vector, resulting from the

concatenation of the three separate vectors, is the new

descriptive representation of the input image.

3.2.2 Features Extraction by Color-ORB

Oriented FAST and Rotated BRIEF (ORB) (El-

Hallak and Lovell, 2013) is a novel approach based on

the original methods Features from Accelerated Seg-

ment Test (FAST) (Rosten et al., 2010) and Binary

Robust Independent Elementary Features (BRIEF)

(Calonder et al., 2010). ORB, as the original works

of FAST, BRIEF and BRISK, is applied to gray-scale

image by calculating the FAST key-points then the

BRIEF descriptors. In this work, we use RGB image

channels separately to compute ORB. The procedure

is the same as Color-BRISK explained above.

3.2.3 Color-BRISK Features and DAISY

Descriptors

In this section, we re-use the Color-BRISK key-

points. These latter are passed as inputs to the Fast

Local Descriptor for Dense Matching (DAISY) (Tola

et al., 2010). We chose this combination to verify if

better results can be obtained by using DAISY de-

scriptor, which is known to be fast and efﬁcient for

Bag-of-Features construction.

3.3 Features Vectors Encoding

In computer vision, descriptors vectors are effective

for image matching, image retrieval or object detec-

tion. For our system, we want to use the descriptors

for classiﬁcation purposes: rather than classifying di-

rectly the image, low-level features are used instead

as its new representations. Raw descriptor vectors can

not be passed directly to the classiﬁcation module; an

encoding on Bag-of-Features (BoF) is necessary.

For descriptor encoding on Bag-of-Features

(BoF), we use the method based on kmeans and the

frequencies histogram. This method is proved to be

efﬁcient in the whole image classiﬁcation task. The

BoFs by kmeans/frequency histogram is performed as

follows:

• After choosing the number of clusters (in our case,

k=5), centroids of each cluster are randomly ini-

tialized by element of the descriptor vectors space.

• For the rest of the descriptor vectors, assign a clus-

ter and recalculate the centroid of the cluster.

A Comparative Study of BRISK, ORB and DAISY Features for Breast Cancer Classiﬁcation

967

• In the end, the descriptor vector are regrouped into

clusters (in our case, 5 groups)

• For the features dictionary (clusters), a histogram

of frequency is assigned. The latter represents

the apparition number of a given descriptor of the

cluster.

3.4 BoF Classiﬁcation Using SVM

First introduced in (Cortes and Vapnik, 1995), Sup-

port Vector Machine (SVM) is a widely used su-

pervised classiﬁcation and regression method. This

method proposed two major contributions to the su-

pervised data classiﬁcation ﬁeld. The ﬁrst aspect of

originality provided by SVM lies in the insurance

of data separability. In fact, the authors state that

if in the original deﬁnition space, the data is inter-

leaved or overlapped, a passage to higher dimensional

space secures the obtaining of a linearly separable re-

deﬁnitions of the original data. The second goal of

SVM is to ﬁnd the optimal linear hyperplane which

ensures the classiﬁcation of the data while optimizing

the margins. In our system, we use the classic binary

SVM. This latter is applied when the experiments data

is contained in twos labelled classes.

4 TESTS AND EVALUATION

4.1 Experimental Setup

The elaboration of our method was performed using a

machine conﬁgured as follows: Intel® Core™ I9 10th

Gen up to 5.30GHz CPU, 32 GB of RAM, NVIDIA®

GeForce® GTX 2080 SUPER GPU, 512 GB SSD.

As a software base, we used the Python 3.8 program-

ming language, the library OpenCV 4.4.0 and its ex-

tra modules for the image pre-processing, features ex-

tracting and their encoding on Bag-of-Features. The

training and evaluation of SVM was fulﬁlled using

TensorFlow CPU 2.4.1.

4.2 Experiments Data

As mentioned before, in the histological process,

there are three possible staining techniques: histo-

chemistry (HC), immunohistochemistry (IHC) and

immunoﬂuorescence (IF), to each method, a multi-

tude of stains can be associated. The most frequently

used ones are: Eosin (E), Hematoxylin (H) and the

combination Eosin-Hematoxylin (H&E). To the bet-

ter of our knowledge, in digital histopathology, there

are no IF databases and a very few IHC databases;

the majority of available data-sets concern HC H&E

breast tumoral slides. Due to this lack of IHC and IF

image bases, we decided to use H&E stained slides

in our experiments. The data-set used is BreakHis

(Spanhol et al., 2016), it offers a collection of H&E

stained SOB slides. The images are categorized,

ﬁrstly, following magniﬁcation and then following the

breast tumor type (benign or malignant). In each cat-

egory, images are grouped following the cells. (See

Tab. 2 for data-set details and total image number for

each category)

Table 2: BreakHis Dataset Details.

Type cell X40 X100 X200 X400

Benign

A 114 113 111 106

F 253 260 264 237

PT 109 121 108 115

TA 149 150 140 130

Malignant

DC 864 903 896 788

LC 156 170 163 137

MC 205 222 196 169

PC 145 142 135 138

The cells categories of the benign class are:

Adenosis (A), Fibroadenoma (F), Phyllodes Tumor

(PT) and Tubular Adenoma (TA). For the malignant

cells, there categories are: Ductal Carcinoma (DC),

Lobular Carcinoma (LC), Mucinous Carcinoma (MC)

and Papillary Carcinoma (PC). The total image con-

tained in the database is equal to 7909 images, 2480

for benign tumor and 5429 for malignant tumor. The

images size is 700 x 420 pixels.

In our work, the image are resized to 201 x 150

pixels. For the system training step, a total of 380

randomly selected images is used, 180 for the benign

class and 180 for the malignant class. In the validation

step, 120 image were used, 60 for benign and 60 for

malignant; these latter were selected randomly from

the original data-set. The purpose of this is to limit

the training data and observe which method obtains

better results.

4.3 Classiﬁcation Results

The interpretation of the extracted features must be

given by a trained machine/deep learning system. As

state before, in this paper, we use Support Vector Ma-

chine to classify our Bag-of-Features (BoF). We eval-

uate our system using two criteria: classiﬁcation ac-

curacy and computation time. The results are shown

in Tab. 3.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

968

Table 3: Experiments results.

Model Classiﬁcation accuracy Precision Recall Computation time

RGB-BRISK/BoF/SVM 72.5% 75.47% 66.67% 4 hours 33 minutes

RGB-ORB/BoF/SVM 65% 68.75% 55% 2 hours 50 minutes

RGB-BRISK/DAISY/BoF/SVM 57.5% 57.89% 55% 4 hours 54 minutes

4.4 Results Discussion

The results of our experiments, as shown in Tab. 3,

prove that even if the learning data is limited, an av-

erage and acceptable classiﬁcation accuracy can be

achieved. From the table, we can retain the follow-

ing:

• The best accuracy value of 72.5% was obtained by

the RGB-BRISK/BoF/SVM system. This system

remains, however, slightly slow.

• The highest value of precision and recall were also

achieved by the RGB-BRISK/BoF/SVM system.

• The fastest system is the one based on RGB-ORB,

however, this latter gave a lower classiﬁcation per-

formances compared to the RGB-BRISK one.

• The RGB-BRISK key-points/DAISY descriptors

method was the slowest and achieved poor classi-

ﬁcation accuracy, precision and recall.

These results were obtained using BoF by kmeans

where K=5, however, we believe that the accuracy

values can increase for a greater k value (10, 20 or

50 and more). In the case of a bigger k value, it is

to be expected that the computation time will drasti-

cally increase. Also, compared to deep learning archi-

tectures, such as ResNet, the RGB-BRISK/BoF/SVM

system remains faster in CPU-based computation. In

the literature works, such as in (Ouddai et al., 2023),

ResNet18 training on similar amount of data took

more than 7 hours. In the case of GPU-based com-

putation, CNN and RNN architectures can be trained

in a signiﬁcantly shorter time.

5 CONCLUSION AND FUTURE

WORKS

In this work, we applied different features extraction

methods for breast tumoral histological slides classi-

ﬁcation. The methods used are: BRISK, ORB and

BRISK/DAISY. An encoding on BoF by kmeans was

performed and the classiﬁcation was done by SVM.

We proposed the exploitation of color information

by computing features from each RGB channel sep-

arately then fusing the three in one. The obtained

results showed that Color-BRISK gave the best clas-

siﬁcation accuracy, Color-ORB was the fastest and

achieved an accuracy of 65% and the combination

Color-BRISK/DAISY gave the worst results in both

computation time and classiﬁcation accuracy. For fu-

ture works, we intend to exploit other hybrid features

extractors, other than BRISK/DAISY. Another per-

spective is to apply BRISK or ORB to other color-

spaces.

REFERENCES

Ai, S., Li, C., Li, X., Jiang, T., Grzegorzek, M., Sun, C., Ra-

haman, M. M., Zhang, J., Yao, Y., and Li, H. (2021).

A State-of-the-Art Review for Gastric Histopathology

Image Analysis Approaches and Future Development.

BioMed Research International, 2021.

Aksac, A., Demetrick, D. J., Ozyer, T., and Alhajj, R.

(2019). BreCaHAD: A dataset for breast cancer

histopathological annotation and diagnosis. BMC Re-

search Notes, 12(1):10–12.

Azevedo Tosta, T. A., Neves, L. A., and do Nascimento,

M. Z. (2017). Segmentation methods of H&E-stained

histological images of lymphoma: A review. Infor-

matics in Medicine Unlocked, 9(May):35–43.

Bukała, A., Cyganek, B., Koziarski, M., Kwolek, B., Ol-

borski, B., Antosz, Z., Swad

zba, J., and Sitkowski,

P. (2020). Classiﬁcation of histopathological images

using scale-invariant feature transform. VISIGRAPP

2020 - Proceedings of the 15th International Joint

Conference on Computer Vision, Imaging and Com-

puter Graphics Theory and Applications, 5(Visigrapp

2020):506–512.

Calonder, M., Lepetit, V., Strecha, C., and Fua, P. (2010).

BRIEF: Binary robust independent elementary fea-

tures. Lecture Notes in Computer Science (including

subseries Lecture Notes in Artiﬁcial Intelligence and

Lecture Notes in Bioinformatics), 6314 LNCS(PART

4):778–792.

Cortes, C. and Vapnik, V. (1995). Support-vector networks.

Machine Learning, 20(3):273–297.

Das, A., Nair, M. S., and Peter, S. D. (2020). Computer-

Aided Histopathological Image Analysis Techniques

for Automated Nuclear Atypia Scoring of Breast Can-

cer: a Review. Journal of Digital Imaging.

El-Hallak, M. and Lovell, D. (2013). ORB an efﬁcient.

Arthritis and Rheumatism, 65(10):2736.

Glotsos, D., Kalatzis, I., Spyridonos, P., Kostopoulos, S.,

Daskalakis, A., Athanasiadis, E., Ravazoula, P., Niki-

foridis, G., and Cavouras, D. (2008). Improving accu-

racy in astrocytomas grading by integrating a robust

least squares mapping driven support vector machine

A Comparative Study of BRISK, ORB and DAISY Features for Breast Cancer Classiﬁcation

969

classiﬁer into a two level grade classiﬁcation scheme.

Computer Methods and Programs in Biomedicine,

90(3):251–261.

Gurcan, M. N., Boucheron, L. E., Can, A., Madabhushi, A.,

Rajpoot, N. M., and Yener, B. (2009). Histopatho-

logical Image Analysis: A Review. IEEE Reviews in

Biomedical Engineering, 2:147–171.

Irshad, H., Jalali, S., Roux, L., Racoceanu, D., Naour, G.,

Hwee, L., and Capron, F. (2013). Automated mito-

sis detection using texture, SIFT features and HMAX

biologically inspired approach. Journal of Pathology

Informatics, 4(2):12.

Irshad, H., Veillard, A., Roux, L., and Racoceanu, D.

(2014). Methods for nuclei detection, segmentation,

and classiﬁcation in digital histopathology: A review-

current status and future potential. IEEE Reviews in

Biomedical Engineering, 7:97–114.

Komura, D. and Ishikawa, S. (2018). Machine Learn-

ing Methods for Histopathological Image Analysis.

Computational and Structural Biotechnology Journal,

16:34–42.

Kostopoulos, S., Glotsos, D., Cavouras, D., Daskalakis,

A., Kalatzis, I., Georgiadis, P., Bougioukos, P., Rava-

zoula, P., and Nikiforidis, G. ANALYTICAL AND

QUANTITATIVE CYTOLOGY AND HISTOLOGY

® ARTICLES Computer-Based Association of the

Texture of Expressed Estrogen Receptor Nuclei with

Histologic Grade Using Immunohistochemically-

Stained Breast Carcinomas. Technical report.

Kumar, M. D., Babaie, M., Zhu, S., Kalra, S., and Tizhoosh,

H. R. (2018). A comparative study of CNN, BoVW

and LBP for classiﬁcation of histopathological im-

ages. 2017 IEEE Symposium Series on Computational

Intelligence, SSCI 2017 - Proceedings, 2018-Janua:1–

Leutenegger, S., Chli, M., and Siegwart, R. Y. (2011).

BRISK: Binary Robust invariant scalable keypoints.

Proceedings of the IEEE International Conference on

Computer Vision, pages 2548–2555.

Li, C., Chen, H., Li, X., Xu, N., Hu, Z., Xue, D., Qi, S.,

Ma, H., Zhang, L., and Sun, H. (2020). A review for

cervical histopathology image analysis using machine

vision approaches, volume 53.

Li, C., Chen, H., Xue, D., Hu, Z., Zhang, L., He, L., Xu, N.,

Qi, S., Ma, H., and Sun, H. (2019a). Weakly Super-

vised Cervical Histopathological Image Classiﬁcation

Using Multilayer Hidden Conditional Random Fields.

In Advances in Intelligent Systems and Computing,

volume 1011, pages 209–221. Springer Verlag.

Li, C., Chen, H., Zhang, L., Xu, N., Xue, D., Hu, Z., Ma,

H., and Sun, H. (2019b). Cervical Histopathology

Image Classiﬁcation Using Multilayer Hidden Condi-

tional Random Fields and Weakly Supervised Learn-

ing. IEEE Access, 7:90378–90397.

Litjens, G., Bandi, P., Bejnordi, B. E., Geessink, O., Balken-

hol, M., Bult, P., Halilovic, A., Hermsen, M., van de

Loo, R., Vogels, R., Manson, Q. F., Stathonikos, N.,

Baidoshvili, A., van Diest, P., Wauters, C., van Dijk,

M., and van der Laak, J. (2018). 1399 H&E-stained

sentinel lymph node sections of breast cancer patients:

The CAMELYON dataset. GigaScience, 7(6):1–8.

Lowe, D. G. (1999). Object recognition from local scale-

invariant features. In Proceedings of the IEEE Inter-

national Conference on Computer Vision, volume 2,

pages 1150–1157. IEEE.

Ninos, K., Kostopoulos, S., Kalatzis, I., Sidiropoulos,

K., Ravazoula, P., Sakellaropoulos, G., Panayiotakis,

G., Economou, G., and Cavouras, D. (2016). Mi-

croscopy image analysis of p63 immunohistochemi-

cally stained laryngeal cancer lesions for predicting

patient 5-year survival. European Archives of Oto-

Rhino-Laryngology, 273(1):159–168.

Ouddai, G., Hamdi, I., and Ben Ghezala, H. (2023). Color-

SIFT Features for Histopathological Image Analysis.

In Lee, R., editor, Software Engineering, Artiﬁcial In-

telligence, Networking and Parallel/Distributed Com-

puting. SNPD 2022. Studies in Computational Intelli-

gence. (in press). Springer cham.

Ozt

urk,

A and Akdemir, B. (2018). Application

of Feature Extraction and Classiﬁcation Methods

for Histopathological Image using GLCM, LBP,

LBGLCM, GLRLM and SFTA. In Procedia Com-

puter Science, volume 132, pages 40–46. Elsevier

B.V.

Popovici, V., Budinsk

a, E.,

apkov

a, L., Schwarz, D.,

sek, L., Feit, J., and Jaggi, R. (2016). Joint analysis

of histopathology image features and gene expression

in breast cancer. BMC Bioinformatics, 17(1):1–9.

Rosten, E., Porter, R., and Drummond, T. (2010). Faster

and better: A machine learning approach to corner de-

tection. IEEE Transactions on Pattern Analysis and

Machine Intelligence, 32(1):105–119.

Sanchez-Morillo, D., Gonz

alez, J., Garc

ıa-Rojo, M., and

Ortega, J. (2018). Classiﬁcation of Breast Cancer

Histopathological Images Using KAZE Features. Lec-

ture Notes in Computer Science (including subseries

Lecture Notes in Artiﬁcial Intelligence and Lecture

Notes in Bioinformatics), 10814 LNBI:276–286.

Shaﬁei, S., Babaie, M., Kalra, S., and Tizhoosh, H. R.

(2021). Colored Kimia Path24 Dataset: Conﬁgura-

tions and Benchmarks with Deep Embeddings.

Simon, O., Yacoub, R., Jain, S., Tomaszewski, J. E., and

Sarder, P. (2018). Multi-radial LBP Features as a Tool

for Rapid Glomerular Detection and Assessment in

Whole Slide Histopathology Images. Scientiﬁc Re-

ports, 8(1):1–11.

Spanhol, F. A., Oliveira, L. S., Petitjean, C., and Heutte,

L. (2016). A Dataset for Breast Cancer Histopatho-

logical Image Classiﬁcation. IEEE Transactions on

Biomedical Engineering, 63(7):1455–1462.

Tola, E., Lepetit, V., and Fua, P. (2010). DAISY: An efﬁ-

cient dense descriptor applied to wide-baseline stereo.

IEEE Transactions on Pattern Analysis and Machine

Intelligence, 32(5):815–830.

Wang, C. W., Chang, C. C., Khalil, M. A., Lin, Y. J., Liou,

Y. A., Hsu, P. C., Lee, Y. C., Wang, C. H., and Chao,

T. K. (2022). Histopathological whole slide image

dataset for classiﬁcation of treatment effectiveness to

ovarian cancer. Scientiﬁc Data, 9(1):1–5.

ICPRAM 2023 - 12th International Conference on Pattern Recognition Applications and Methods

970