Saliency Maps of Video-colonoscopy Images for the Analysis of Their

Content and the Prevention of Colorectal Cancer Risks

Valentine Wargnier-Dauchelle

, Camille Simon-Chane

and Aymeric Histace

ETIS Lab, ENSEA, 6 Av. du Ponceau, Cergy-Pontoise, France

Keywords:

Colorectal Cancer, Video-colonoscopy, Polyp Detection, Computer Aided Diagnosis, Saliency Maps, Bag of

Features, SVM, Approximate Entropy.

Abstract:

The detection and removal of adenomatous polyps via colonoscopy is the gold standard for the prevention of

colon cancer. Indeed, polyps are at the origins of colorectal cancer which is one of the deadliest diseases in

the world. This article aims to contribute to the wide range of methods already developed for the prevention

of colorectal cancer risks. For this, the work is organized around the detection and the localization of polyps

in video-colonoscopy images. The aim of this paper is to ﬁnd the best description of a bowel image in order

to classify a patch, that is to say a image fragment, as polyp or not. The classiﬁcation is achieved thanks

to an SVM (Support Vector Machine) using a bag of features. Different types of features extraction will be

compared. Thus, the traditional SURF (Speeded-Up Robust Features) extractor will be compared to local

features extractors like HOG (Histogram of Oriented Gradient) and LBP (Local Binary Pattern) but also to an

original extractor based on the structural entropy.

1 INTRODUCTION

In France, bowel cancer is the second most common

cancer in women and the third most common in men.

In 2017, 45 000 new cases were reported, and col-

orectal cancer caused the death of 9294 men and 8390

women in France. Survival rates vary greatly based

on the progress of the cancer when it is detected and

the beginning of the medical treatment (surgical or

drugs). Indeed, the chances of survival are around

90% if the cancer is detected during the ﬁrst stage

whereas they are around only 5% for stage V.

For these reasons, the early detection of polyps

is fundamental. However, studies show that 26%

of polyps present in the gut remain undetected by

doctors during a video-colonoscopy. Some are sim-

ply invisible to the camera, others are present in the

video stream and therefore detectable. Thus, com-

puter aided diagnosis is a major issue in the diagnosis

of colorectal cancer. Many methods have been pro-

posed in recent years to reduce polyp miss rate and

improve detection capabilities. These methods can be

divided into three groups : ad-hoc, machine learning

and hybrid methods.

https://orcid.org/0000-0003-1883-791X

https://orcid.org/0000-0002-4833-6190

https://orcid.org/0000-0002-3029-4412

The majority of ad-hoc methods are based on ex-

ploiting low-level image processing methods to esti-

mate candidate polyp boundaries. For example, (Iwa-

hori et al., 2013) use Hessian ﬁlters, (Bernal et al.,

2015), intensity valleys and (Silva et al., 2014), the

Hough transform. The extracted information is then

used to localize polyps using the curvatures analysis

in the work of (Zhu et al., 2010), the ellipsoidal shape

search according to the (Kang and Doraiswami, 2003)

or a combination of both for (Hwang et al., 2007).

For machine learning methods, texture and color

information were often used as descriptors such as

color wavelets in the work of (Karkanis et al., 2003),

coocurrence matrices for (Ameling et al., 2009) or lo-

cal binary patterns (LBP) which are exploited in the

work of (Gross et al., 2009). Some of the most recent

methods use deep learning as in the work of (Ribeiro

et al., 2016) among many others in the last four years.

Finally, hybrid methods combine both methodolo-

gies for polyp detection. For instance, (Tajbakhsh

et al., 2014) combine edge detection and feature ex-

traction, (Silva et al., 2014) use hand-crafted features

to ﬁlter non-informative image regions and (Ševo

et al., 2016) mix edge density and convolutional net-

works.

The performances of these methods can be evalu-

ated on two criteria: accuracy and computation time.

106

Wargnier-Dauchelle, V., Simon-Chane, C. and Histace, A.

Saliency Maps of Video-colonoscopy Images for the Analysis of Their Content and the Prevention of Colorectal Cancer Risks.

DOI: 10.5220/0009148401060114

In Proceedings of the 13th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2020) - Volume 4: BIOSIGNALS, pages 106-114

ISBN: 978-989-758-398-8; ISSN: 2184-4305

Indeed, a good polyp detection algorithm must detect

most of the polyps without too many false alarms.

To be clinically applicable, it also has to meet real-

time constraints. For videos acquired at 25 frames

per second that corresponds to a maximal processing

time of 40 ms per frame. A comparison of recent

methods performances is achieved in (Bernal et al.,

2017). Ad-hoc method often present good computa-

tion time but a weaker accuracy and conversely for

machine learning methods. Active learning method-

ologies have been introduced in the work of (Anger-

mann et al., 2016) to reinforce the compromise be-

tween performance and computation time.

The recent YOLOv3 deep learning architec-

ture (Redmon and Farhadi, 2018) represents a ma-

jor step further for reliable and real-time polyp de-

tection. Nevertheless, the need for speciﬁc computa-

tion resources (GPU), can be seen as a limitation for a

routine to be used and integrated in standard colono-

scope.

Alternatively to "fast" deep learning approach like

Yolo, an improved polyp detector, in terms of sensi-

tivity and speciﬁcity, can be designed by deﬁning a

saliency detection approach than can be both used for

direct detection or for the reduction of false alarms

using classic shallow methods. In this paper, as an al-

ternative to deep learning approaches, we propose a

saliency-based strategy which aim is to compare dif-

ferent types of classic but also original feature extrac-

tors in order to ﬁnd the more relevant for polyp detec-

tion and localization tasks.

Our method is based on the previous work

of (Raynaud et al., 2019) about localization of an-

giodysplasias in videocapsule images. They devel-

oped an active contour segmentation approach for

small bowel lesions characterization using saliency

maps as extractors. These saliency maps are gen-

erated using a dictionary learning strategy (bag of

words) associated with a binary SVM classiﬁer. More

precisely, the classiﬁer allows for a given input image

to create a probability map of angiodisplasia presence

using a sliding window of predeﬁned size all over the

image for which center-pixel is associated to the prob-

ability given by the SVM classiﬁer. We propose here

to investigate this method for polyp detection and lo-

calization.

In this paper, we propose to test and estimate per-

formance of different features extractors that will be

used for the dictionary learning step, including classic

ones from the literature, such as SIFT (Scale-Invariant

Features Transform) but also an original one based on

approximate entropy (Histace et al., 2014).

The paper is organized as follows: in section 2 we

introduce our methodology and present the different

(a) (b)

Figure 1: Examples of patches used for training: (a) Polyp

patches. (b) Non-polyp patches.

features extractor used. Then section 3 presents the

experiments and the results for each extractor. Finally,

section 4 concludes the paper.

2 METHODOLOGY

Our method can be divided into two parts. First, we

have achieved the training and the evaluation of the

different bag of features using patches extracted from

the classic set of colonoscopic data known as CVC-

Clinic and CVC-Colon database (Bernal et al., 2017).

This training aims to learn how to classify a patch as

a polyp or not from features extracted by the bag of

words. The evaluation on patches permits to judge the

capacity of a two-class SVM, based on a speciﬁc bag

of words, to correctly classify a given set of patches.

Based on this, for a given image (taken from a differ-

ent set of the videocolonoscopic images as those used

for training), a saliency map is generated and an eval-

uation of obtained saliency map algorithm is achieved

(section 2.2). The aim of this evaluation is to assess

how the different descriptors perform on complete im-

ages to get a ﬁrst estimation of the performance re-

lated to a given feature extractor.

2.1 Training

In a polyp detection context, we have tested several

type of features in order to ﬁnd the best description of

images for colonoscopy. The ﬁrst step consists in gen-

erating all the descriptors for the images, that is to say,

ﬁnding the interest points and describing interest ar-

eas around interest points that will be used for the bag

of features dictionary. The next step is to generate the

associated optimal dictionary. For this, we classically

used the k-means clustering over the descriptors to

obtain the representative features of the images which

will be the words of the dictionary. These words will

be used by the SVM for the classiﬁcation.

The training is done on patches (see examples

Figure 1). All images considered in this paper are

from the CVC Colon and CVC Clinic databases.

Our speciﬁc patches database was composed of 4412

Saliency Maps of Video-colonoscopy Images for the Analysis of Their Content and the Prevention of Colorectal Cancer Risks

107

Table 1: Confusion matrix for the SURF descriptor.

PREDICTED

KNOWN Negative Positive

Negative 0.98 0.02

Positive 0.06 0.94

negative patches and 942 positive patches from the

two databases. 80% of them were used for training

and 20% for the evaluation, ﬁrst made on a patches

classiﬁcation task. Negative patches were randomly

selected from the negative images whereas positive

patches were generated from the pixel-wise polyp

ground truth. Three images and there associated

patches were kept out of these databases to generate

the saliency map and evaluate the method on complete

images.

The training and the evaluation are repeated three

times on different set of patches for each of these pa-

rameters combinations: the extraction of patches be-

ing random, the training and the evaluation database

were different for each test.

The patch classiﬁcation is evaluated using the con-

fusion matrix which gives the percentage of true pos-

itives, false negatives, true negatives and false posi-

tives. For the second step, that is the saliency map de-

tection approach, a psychovisual metric is proposed.

In the following section, short descriptions of the

considered feature descriptors are given along with

obtained results for the patch classiﬁcation tasks.

2.1.1 SURF Descriptor

The SURF descriptor is considered for its improved

performances in terms of robustness and speed com-

pared to the SIFT descriptor. For the purpose of hav-

ing the best bag of features, we have tested several

combination of parameters namely the dictionary size

(from 500 to 4000 visuals words) and the detector

(SURF and square grid with a size varying from 4 x 4

to 12 x 12 pixels). The SVM kernel has also been cho-

sen in accordance with the best obtained performance.

We have tested the Gaussian, the linear and the poly-

nomial (order 2 and 4) kernel. Finally, we have tested

different types of input images: classical RGB image,

but also each component taken separately (red, green

and blue channels).

For this descriptor the best classiﬁer used a dictio-

nary with 800 words, a square grid (4 x 4 pixels) de-

tector, a multi-scale SURF descriptor, a polynomial

kernel (order 2) and the blue channel as input. The

performance results for this classiﬁer are presented in

the confusion matrix of Table 1.

The blue channel is often used in polyp detection

because it is the most representative of the polyp in-

Table 2: Confusion matrix for the SURF + LBP descriptor.

PREDICTED

KNOWN Negative Positive

Negative 0.98 0.02

Positive 0.11 0.89

formation, as shown by (Bernal et al., 2015). Indeed,

they prove that the blue channel permits to mitigate

useless information like blood vessels in a context of

valley detection. But it can be generalized for all

polyp detection approaches. The size of the dictio-

nary, the detector and the kernel will be the same for

all the bags of words tested.

2.1.2 LBP Descriptor

We have decided to use a LBP descriptor in addi-

tion to the SURF descriptor. Indeed, this type of de-

scriptor adds another local information. Moreover, it

has given satisfying results in the study of (Anger-

mann et al., 2016). This can be explained by the fact

that LBP descriptors code texture information and it

has been proven that polyps have a typical texture as

shown in (Ameling et al., 2009).

This is evaluated on the blue channel, as used

in (Angermann et al., 2016) and because it was the

most relevant channel for the SURF descriptor. The

SURF descriptor is the same as in the previous expe-

rience.

Using the same evaluation protocol, we obtain

the confusion matrix presented in Table 2. With an

unchanged True-Negative detection rate, the True-

Positive detection rate is slightly lower than for the

SURF descriptor alone.

2.1.3 HOG Descriptor

Another classic local descriptor is the HOG descrip-

tor. This type of descriptor is especially efﬁcient for

edge detection. (Iwahori et al., 2013) use this type of

descriptor in their polyp detection method with Hes-

sian ﬁlters because polyp and non-polyp regions have

approximately the same color which is not taken into

account by the HOG descriptor.

The HOG features are extracted from the same

multi-scale grid as the SURF descriptor. Because

HOG features are very dense, we have decided to

use it alone, without SURF features. This descrip-

tor was tested on both grayscale and blue channel im-

ages. Always with the same validation conditions,

we have obtained the confusion matrices given in Ta-

bles 3 and 4. In this case there is a slight better True-

Positive detection rate on the grayscale images com-

pared to the blue channel.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

108

Table 3: Confusion matrix for the HOG descriptor with blue

channel.

PREDICTED

KNOWN Negative Positive

Negative 0.97 0.03

Positive 0.19 0.81

Table 4: Confusion matrix for the HOG descriptor with

grayscale images.

PREDICTED

KNOWN Negative Positive

Negative 0.97 0.03

Positive 0.17 0.83

2.1.4 Classic Descriptors Applied on

Approximate Entropy Maps

In this section, we propose to consider an original ap-

proach for indirect feature extraction, based on Ap-

proximate Entropy (ApEn) formerly introduced for

active contour image segmentation by (Histace et al.,

2014). The ApEn, also called structured entropy, is

a statistical metric which measures the regularity in a

sequence of numerical data. It measures the proba-

bility for two segments, which are extracted from the

same sequence, to stay close if their length is incre-

mented by one.

The parameters are the length of the sequence N,

the length of the extracted segments m and the ﬁlter-

ing level r which imposes the necessary similarity be-

tween two sequence to be considered as close. For in-

stance, if we consider the following sequence of num-

bers:

u = [u(1), u(2) ... u(k), u(k + 1) ... u(k + m) ... u(N)]

(1)

We can now construct a series x(i) with N − m + 1

segments from u and where x(i) = [u(i), ...u(i + m)] is

a segment of length m.

Then, we can calculate the coefﬁcient :

(r) =

Number of x( j) such that d(x(i), x( j)) ≤ r

N − m + 1

(2)

where:

d(x(i), x( j)) =

∑

k=1

(i) − x

( j)| (3)

The average of the coefﬁcients is :

(r) =

N − m + 1

N−m+1

∑

i=1

(r) (4)

(a)

(b) (c)

Figure 2: Approximate entropy images: (a) Original image.

(b) ApEn on grayscale image. (c) ApEn on blue channel

image.

All of this permits to deﬁne the approximate en-

tropy as :

ApEn = ln

(r)

m+1

(r)

(5)

Because it measures the regularity in a sequence,

it is efﬁcient to determine homogeneity changes as

Pincus has proven in 1901. For this reason, the metric

can be used to detect edges and in our case, polyps.

(Nagy et al., 2019) use this entropy in order to de-

tect polyps thanks to their speciﬁc curves. However,

this entropy has never been used in a machine learn-

ing context. We propose to use it in polyp detection

with this approach.

For this, we have chosen a ﬁltering level r = 1.75,

a sequence length N = 9 × 9 corresponding to a vec-

torised thumbnail (9 × 9 pixels) and a segment length

m = 2 as (Histace et al., 2014) chose. For illustra-

tion, we have computed the distance map in term of

approximate entropy for video-colonoscopy images.

We have applied this on grayscale and blue channel

images. The results are presented on Figure 2.

We show that approximate entropy is a good edge

detector especially when it is applied on grayscale im-

ages. We propose to use this type of image as input

image for our classiﬁcation thanks to a SURF descrip-

tor which is the most used in features extraction and

a HOG descriptor. Indeed, the HOG descriptor is par-

ticularly appropriate for edges and the approximate

entropy map emphasizes them. The confusion matri-

ces are given in Tables 5 and 6.

Saliency Maps of Video-colonoscopy Images for the Analysis of Their Content and the Prevention of Colorectal Cancer Risks

109

Figure 3: Proposed method. A variant consists in replacing the blue channel by a grayscale image or a distance map.

Table 5: Confusion matrix for the SURF descriptor applied

on ApEn images.

PREDICTED

KNOWN Negative Positive

Negative 0.97 0.03

Positive 0.07 0.93

Table 6: Confusion matrix for the HOG descriptor applied

on ApEn images.

PREDICTED

KNOWN Negative Positive

Negative 0.95 0.05

Positive 0.26 0.74

2.2 Saliency Map Generation

In order to evaluate the capacities of our descriptors

in a clinical context, that is to say on a complete im-

age, we have used the method introduced in (Raynaud

et al., 2019). The saliency map algorithm takes an

image as input (see Figure 3). Only the blue chan-

nel is used for most of our descriptors but grayscales

images and approximate entropy maps are also used.

Then the image is divided in patches shifted by one

pixel in the two direction of the space, thanks to a

moving windows. We have chosen a 100 × 100 pix-

els window size, corresponding to the average size

of polyps in the database. For each patch, features

are extracted according to the bag of features created

during the training. These features permit to feed the

SVM in order to classify the patch and to obtain a re-

lated probability classiﬁcation. Finally, a heat map is

created from the distance to the positive class for each

patch associated to its central pixel.

3 EXPERIMENTS AND RESULTS

To evaluate performance of the different feature ex-

tractors previously introduced we test the method on

three speciﬁc images (ﬁrst row of Figure 4). These

images were selected from the CVC-Clinic database

for their characteristics, making them representative

of the different types of polyps that can be found in

a clinical context: a ﬂat polyp seen from above, a

pedunculated polyp seen from the side and a polyp

which is slightly hidden by bowel folds. Then we can

create the probability maps with the different descrip-

tors, through the SVM local response as described in

previous section. These results are shown in Figure 4.

We ﬁrst notice that all the descriptors predict the

specular highlight as a polyp. Indeed, the third image

presents very clearly a specular light spot at the bot-

tom left and the polyp presence probability is high in

this area for all the descriptors. It is due to the typ-

ical form of a polyp, called blob, which predisposes

light reﬂection. The only exception is the SURF de-

scriptor using the approximated entropy (Figure 4(r)).

It is not sensitive to this type of reﬂection which is a

major advantage since specular light is a problem in

many polyp detection methods of the literature.

Additionally, results show that the LBP feature in

addition of the SURF descriptor (Figure 4(g, h, i)),

permit to reinforce the decision taken with the SURF

descriptor only (Figure 4(d, e, f)).

The HOG descriptor with grayscale images (Fig-

ure 4(m, n, o)) is not at all speciﬁc as almost the entire

image is considered as a polyp. On the other hand,

the HOG descriptor applied on the blue channel (Fig-

ure 4(j, k, l)) of the images is very speciﬁc. Indeed,

the polyp presence probability higher than 0,5 only on

the polyp, especially for the ﬂat polyp (Figure 4(j)).

The HOG descriptor using the approximate entropy

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

110

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

(j) (k) (l)

Figure 4: Saliency maps where the ground truth is drawn in white: (a, b, c) Original image. (d, e, f) With SURF descriptor.

(g, h, i) With SURF and LBP descriptor. (j, k, l) With HOG descriptor on blue channel. (m, n, o) With HOG descriptor on

grayscale images. (p, q, r) With SURF descriptor on ApEn images. (s, t, u) With HOG descriptor on ApEn images.

Saliency Maps of Video-colonoscopy Images for the Analysis of Their Content and the Prevention of Colorectal Cancer Risks

111

(m) (n) (o)

(p) (q) (r)

(s) (t) (u)

Figure 4: Saliency maps where the ground truth is drawn in white: (a, b, c) Original image. (d, e, f) With SURF descriptor.

(g, h, i) With SURF and LBP descriptor. (j, k, l) With HOG descriptor on blue channel. (m, n, o) With HOG descriptor on

grayscale images. (p, q, r) With SURF descriptor on ApEn images. (s, t, u) With HOG descriptor on ApEn images (cont.).

(Figure 4(s, t, u)) also presents this feature. The prob-

ability is higher on the polyp with the second but it is

a little bit less speciﬁc. This high speciﬁcity could be

very interesting in polyp detection.

In fact, all the current methods have a very good

sensitivity but weak speciﬁcity. It means that almost

all the polyps are detected but the method introduces a

lot of false alarms. This is a problem because they can

distract the doctor during the colonoscopy. Moreover,

ﬂat polyps are a major challenge of computer aided

diagnosis for colonoscopy because most of the unde-

tected polyps during this exam are ﬂat polyps. Polyps

like the second and third images are easily detectable

by the clinician. Thereby, our method can be use as

a reﬁnement method pre-processed colonoscopy im-

ages. The pre-processing could be performed by a

method among those proposed in the literature. Our

reﬁnement method could then be applied on the parts

of the image deﬁned as polyps by the ﬁrst method.

This step would eliminate false alarms, improving the

performances of current methods.

4 CONCLUSION AND

DISCUSSION

In this paper, we compared several colonoscopy im-

age descriptors for bag of words. We show that the

HOG descriptor applied on the blue channel of the

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

112

(a) (b)

(e) (f)

Figure 5: Tests on patches with the SURF descriptor: (a, b)

Patch without polyp and corresponding map. (c, d) Patch

with a easily detectable polyp and corresponding map. (e,

f) Patch with a less visible polyp and corresponding map.

(a) (b) (c)

Figure 6: Patches size test with the SURF descriptor: (a)

Original image. (b) Saliency map with a 100 x 100 pixels

window. (c) Saliency map with a 150 x 150 pixels window.

image could be use as a conﬁrmation in another polyp

detection method. We point also the insensitivity to

specular light of the SURF descriptor associated to

the approximated entropy. As the specular light is

a problem in most of the polyp detection method, it

could be useful associated to another method. An-

other solution is to use images where specular light

has been removed by image processing as in (Sánchez

et al., 2017).

These results also show that the detection spreads

around the ground truth. A solution could be to re-

inforce our training database adding offset patches.

Indeed, our current database only present centered

polyps. For more robustness and to train our algo-

rithm to detect cut or decentered polyps, the diversiﬁ-

cation of the database is necessary.

At last, the confusion matrix and the tests on

patches are good as shown in the Figure 5. Never-

theless, the tests on complete images are not totally

satisfactory. This could be partly due to the weak ro-

bustness of the database but it is also due to the size

of the moving window in the segmentation algorithm.

This parameter is fundamental for the success of our

method. Our work is based on the comparison of im-

ages descriptors with stable settings but, in order to

use the method for polyp detection, this parameter

must be adjusted. Indeed, the results are completely

different according when varying the moving window

size, as shown in the Figure 6.

The study proposed here shows that our method

opens realistic alternative to CNN (Convolutional

Neural Network) approaches, even if the parameter

settings needs to be improved in order to optimize the

saliency map generation. This latter could then be-

come, on its own, a detection map, but also used to

reduce the False-Positive rate related to classic ma-

chine learning methods such as boosting or SVM for

instance.

REFERENCES

Ameling, S., Wirth, S., Paulus, D., Lacey, G., and Vi-

larino, F. (2009). Texture-based polyp detection in

colonoscopy. In Bildverarbeitung für die Medizin

2009, pages 346–350. Springer.

Angermann, Q., Histace, A., and Romain, O. (2016). Ac-

tive learning for real time detection of polyps in video-

colonoscopy. Procedia Computer Science, 90:182–

187.

Bernal, J., Sánchez, F. J., Fernández-Esparrach, G., Gil,

D., Rodríguez, C., and Vilariño, F. (2015). Wm-dova

maps for accurate polyp highlighting in colonoscopy:

Validation vs. saliency maps from physicians. Com-

puterized Medical Imaging and Graphics, 43:99–111.

Bernal, J., Tajkbaksh, N., Sánchez, F. J., Matuszewski, B. J.,

Chen, H., Yu, L., Angermann, Q., Romain, O., Rus-

tad, B., Balasingham, I., et al. (2017). Compara-

tive validation of polyp detection methods in video

colonoscopy: results from the miccai 2015 endo-

scopic vision challenge. IEEE transactions on med-

ical imaging, 36(6):1231–1249.

Gross, S., Stehle, T., Behrens, A., Auer, R., Aach, T., Wino-

grad, R., Trautwein, C., and Tischendorf, J. (2009). A

comparison of blood vessel features and local binary

patterns for colorectal polyp classiﬁcation. In Medical

Imaging 2009: Computer-Aided Diagnosis, volume

Saliency Maps of Video-colonoscopy Images for the Analysis of Their Content and the Prevention of Colorectal Cancer Risks

113

7260, page 72602Q. International Society for Optics

and Photonics.

Histace, A., Bonnefoye, E., Garrido, L., Matuszewski, B. J.,

and Murphy, M. F. (2014). Active contour segmen-

tation based on approximate entropy - application to

cell membrane segmentation in confocal microscopy.

In BIOSIGNALS 2014 - Proceedings of the Interna-

tional Conference on Bio-inspired Systems and Signal

Processing, ESEO, Angers, Loire Valley, France, 3-6

March, 2014, pages 270–277.

Hwang, S., Oh, J., Tavanapong, W., Wong, J., and

De Groen, P. C. (2007). Polyp detection in

colonoscopy video using elliptical shape feature. In

Image Processing, 2007. ICIP 2007. IEEE Interna-

tional Conference on, volume 2, pages II–465. IEEE.

Iwahori, Y., Shinohara, T., Hattori, A., Woodham, R. J.,

Fukui, S., Bhuyan, M. K., and Kasugai, K. (2013).

Automatic polyp detection in endoscope images using

a hessian ﬁlter. In proceedings of MVA conference,

pages 21–24.

Kang, J. and Doraiswami, R. (2003). Real-time image

processing system for endoscopic applications. In

CCECE 2003-Canadian Conference on Electrical and

Computer Engineering. Toward a Caring and Hu-

mane Technology (Cat. No. 03CH37436), volume 3,

pages 1469–1472. IEEE.

Karkanis, S. A., Iakovidis, D. K., Maroulis, D. E., Karras,

D. A., and Tzivras, M. (2003). Computer-aided tumor

detection in endoscopic video using color wavelet fea-

tures. IEEE transactions on information technology in

biomedicine, 7(3):141–152.

Nagy, S., Sziová, B., and Pipek, J. (2019). On structural en-

tropy and spatial ﬁlling factor analysis of colonoscopy

pictures. Entropy, 21(3):256.

Raynaud, G., Simon-Chane, C., Jacob, P., and Histace, A.

(2019). Active contour segmentation based on his-

tograms and dictionary learning. In proceedings of

VISAPP conference, pages 609–615.

Redmon, J. and Farhadi, A. (2018). Yolov3: An incremental

improvement. arXiv.

Ribeiro, E., Uhl, A., and Häfner, M. (2016). Colonic

polyp classiﬁcation with convolutional neural net-

works. In Computer-Based Medical Systems (CBMS),

2016 IEEE 29th International Symposium on, pages

253–258. IEEE.

Sánchez, F. J., Bernal, J., Sánchez-Montes, C., de Miguel,

C. R., and Fernández-Esparrach, G. (2017). Bright

spot regions segmentation and classiﬁcation for spec-

ular highlights detection in colonoscopy videos. Ma-

chine Vision and Applications, 28(8):917–936.

Ševo, I., Avramovi

c, A., Balasingham, I., Elle, O. J., Bergs-

land, J., and Aabakken, L. (2016). Edge density based

automatic detection of inﬂammation in colonoscopy

videos. Computers in biology and medicine, 72:138–

150.

Silva, J., Histace, A., Romain, O., Dray, X., and Granado,

B. (2014). Toward embedded detection of polyps in

wce images for early diagnosis of colorectal cancer.

International Journal of Computer Assisted Radiology

and Surgery, 9(2):283–293.

Tajbakhsh, N., Gurudu, S. R., and Liang, J. (2014). Au-

tomatic polyp detection using global geometric con-

straints and local intensity variation patterns. In In-

ternational Conference on Medical Image Computing

and Computer-Assisted Intervention, pages 179–187.

Springer.

Zhu, H., Fan, Y., and Liang, Z. (2010). Improved curva-

ture estimation for shape analysis in computer-aided

detection of colonic polyps. In International MICCAI

Workshop on Computational Challenges and Clinical

Opportunities in Virtual Colonoscopy and Abdominal

Imaging, pages 9–14. Springer.

BIOSIGNALS 2020 - 13th International Conference on Bio-inspired Systems and Signal Processing

114