QUALITY ASSESSMENT IN COLONOSCOPY

New Challenges Through Computer Vision-based Systems

Fernando Vilari˜no

Computer Vision Center and Computer Science Dep. Universitat Aut`onoma de Barcelona, Spain

Gerard Lacey

Graphics, Vision and Visualisation Group (GV2), School of Computer Science and Statistics, Trinity College Dublin, Ireland

Keywords:

Colonoscopy, Quality assessment, Eye-tracker, Computer vision, Polyps, Colon cancer.

Abstract:

The assessment of the quality of the colonoscopic interventions arises as a most relevant issue once the number

and the availability of these clinical procedures are increased day by day. The use of the latest computer vision-

based techniques can provide the physician with both qualitative and, most important, objectively veriﬁable

quantitative indicators of performance. In this paper we present a study in which we propose the automatic

analysis of colonoscopy video for the quality assessment of the intervention from different points of view:

1) We propose the characterization of the different parts of the colon in order to obtain metrics of the time

used for navigation, portion of gut analyzed, etc. 2) We analyze the image contents in order to automatically

characterize the presence of polyps. 3) We use the information obtained by and eye-tracker in order to assess

the physician’s skills.

1 COLON CANCER AND

INTESTINAL SCREENING

1.1 Colon Cancer in Numbers

The main lesions associated to the intestine are:

bleeding, lump, ulcer, Crohn disease, and cancer.

During the last 20 years, colon cancer has been

the second leading cause of cancer deaths in the

United States, behind lung cancer, with approxi-

mately 60, 000 deaths per year as shown in O’Brien’s

report (O’Brien et al., 1990), and also analyzed in

other studies (U.S. Department of Health and Human

Services, 2003). America Cancer Society’s 2007 re-

port (American Cancer Society, 2007) provides an ex-

tended summary of facts and statistics about colon

cancer prevalence and impact in the population. Col-

orectal cancer is the second leading cause of cancer-

related deaths in Singapore and Europe (Ministry of

Health Singapore, 1990). Although colon screening

has become the main alternative for prevention of col-

orectal cancer, recent data suggest that there is a sig-

niﬁcant miss-rate for the detection of even relatively

large polyps and cancer (Pabby et al., 2005). For this

reason, special efforts havebeen focused on the devel-

opment of computer-aided systems for the detection

of this type of pathologies. Nowadays, novel lines

of research are oriented to widen this perspective by

means of the implementation of objective indicators

for the assessment of the procedures used in colon

screening, since the miss rate of polyps is highly cor-

related to elements such as the quality of the prepa-

ration, the time consumed for the intervention, the

amount of intestinal surface screened, the kills of the

physician in the manipulation of the endoscope, etc.

Moreover, this indicators present a potential value for

the training of future endoscopists as reference values

to measure abilities in and objective way.

1.2 Screening Technics

Intestinal endoscopy is referred to as the technique

for screening the intestinal lumen. For the case

of large intestine, this technique receives the name

of colonoscopy. Fiberoptic colonoscopy (FOC) is

widely accepted as the deﬁnitive method for diagno-

sis of colonic polyps. FOC allows direct visualiza-

tion of the intestinal surface and affords the possibil-

ity of obtaining a in-situ biopsy as well as cauteriza-

320

Vilariño F. and Lacey G. (2009).

QUALITY ASSESSMENT IN COLONOSCOPY - New Challenges Through Computer Vision-based Systems.

In Proceedings of the International Conference on Biomedical Electronics and Devices, pages 320-325

DOI: 10.5220/0001780703200325

 SciTePress

tion and clinical intervention such as polyp removal

(Winawer et al., 1997). FOC is a minimal invasive

surgery (Hunter and Sackier, 1993), (Hulka and Re-

ich, 1994) consisting of the introduction through the

anus of a ﬂexible probe which a camera and an il-

lumination device on its tip. The probe consists of

a ﬂexible cable which can be controlled by the ex-

pert in order to reach every part on the intestinal wall.

There exist several reference books regarding clini-

cal colonoscopy for the reader interested in deepen-

ing in the speciﬁcities of colonoscopy, amongst which

Kato’s textbook (Kato and Baron, 2003) can provide

an insightful introductory view. Several authors have

assessed that endoscopic images possess rich infor-

mation (Nagasako et al., 1998), which facilitates the

abnormality detection by multiple techniques (Zheng

et al., 2005). The main drawbacks related with FOC

can be enumerated as follows (Winawer et al., 1993),

(Eddy, 1990): risks of perforation; costs of the inter-

vention; difﬁculty in visualizing the 100% of the in-

testinal surface; high number of patients for a reduced

number of specialists provide a stress and shrink in

the intervention time; difﬁcult visualization due to

the intestinal content; preparation needed; imprecise

localization of events for a subsequent interventions,

and high rate of negative results in inspections look-

ing for polyps (Gokturk and Tomasi, 2001).

In the last years, other modalities for the visual-

ization of the colon have arisen. One of the most

relevant is virtual colonoscopy, a technique consist-

ing of the construction of a virtual 3D model of the

colon from computed tomography (CT) data (Liang

et al., 2004), and for this reason several authors have

proposed automatic techniques for the detection of le-

sions in this modality of images. On the other hand,

Wireless Capsule Video Endoscopy (WCVE) (Iddan

and Meron, 2000) has devoted particular attention, re-

cently. WCVE (Fireman et al., 2002) consists of a

capsule with a camera, a battery and a set of led lamps

for illumination attached to it, which is swallowed by

the patient, emitting a radio frequency signal which is

received and stored in an external device. The result

is a video movie which records the trip of the capsule

along the intestinal tract with a rate of two frames

per second, and that can be easily downloaded into

a PC with the camera software installed. It is much

less invasive, since the patient simply has to swallow

the pill, which will be expelled in the normal cycle

through defecation. Moreover, there is no need of

hospitalization nor expert support through the process

and the patient can lead an ordinary life, since the at-

tached device is recording the video movie emitted

by the camera in the capsule (Vilari˜no, 2006; Vilari˜no

et al., 2009). However, although these new modalities

provide new ways of clinical analysis, colonoscopy is

the reference technique for clinical intervention in the

case of colon screening and colon cancer.

2 COMPUTER VISION FOR

QUALITY ASSESSMENT

We state that computer vision-based techniques can

be used in order to obtain objective indicators of the

quality of the interventions in an automatic way. Our

approach provides a framework for the quality assess-

ment which is implemented in the following 3 areas

in which we are developing our research, namely: 1)

Computer-aided intervention: On-line detection and

characterization of potential targets in intervention-

time, 2) Post-interventional quality metrics: Auto-

matic computation of the quality measures related

to the intervention such as quality of preparation,

amount of bowel surface visualized, time measures,

etc. 3) Evaluation of skills: Analysis of 3D trajec-

tories of the endoscopes and the screening behavior.

For this ﬁnal point, we propose to use the trajecto-

ries together with the information obtained from the

tracking of the gaze position of trainees in order to

assess their skills. The perspective presented in this

paragraph is graphically depicted in Figure 1.

Figure 1: Our framework for automatic quality assessment

in colonoscopy.

3 COLON CANCER: CLINICAL

CHARACTERIZATION

Colon cancer has devoted wide attention in many

studies in order to ﬁnd proper descriptions and cat-

egorization (Rembacken et al., 2000), (Saitoh et al.,

2001), (Paris Workshop Participants, 2003). Adeno-

matous polyps, particularly those larger than 1 cm in

QUALITY ASSESSMENT IN COLONOSCOPY - New Challenges Through Computer Vision-based Systems

321

diameter, are the most likely precursors of colorectal

carcinoma (Gokturk and Tomasi, 2001), (Thoeni and

Laufer, 1994). The main features used for cancer and

general abnormality characterization are: color, shape

and texture.

• Color. Color colonoscopic images tend to ex-

hibit the same color features for the same colon

status (Kato and Baron, 2003). Malignant tu-

mors are usually inﬂated and inﬂamed and this in-

ﬂammation is usually reddish and more severe in

color than the surrounding tissues. Benign tumors

exhibit less intense hues. Redness may specify

bleeding and black may be treated as deposits due

to laxatives. Green may be the presence of fae-

cal materials, which are not clear during the pre-

operative preparation, and yellow relates to pus

formation (Tjoa and Krishnan, 2003).

• Shape. Shape is a relevant cue since polyps are

associated to rounded or peduncular shapes. Pe-

duncular polyps are relatively easy to visualize

during a screening session. Flat polyps present

a higher difﬁculty, and in addition, they are more

likely to develop into malignant polyps.

• Texture. Texture is known to be an important

cue to be evaluated for the discrimination between

malignant and benign lesions (Kudo and Kashida,

2000), (Nagata et al., 2000).

The high-level characterization of cancer ex-

plained above is usually translated into an image-

based feature extraction stage, focused on color, tex-

ture or shape cues. Following the feature extraction,

a discrimination procedure, based on simple compar-

isons or more sophisticated machine learning tech-

niques must be applied. In our approach, efﬁcient

methods for color, texture and shape characteriza-

tion must be oriented towards the high speed require-

ments of on-line procedures. The use of histogram

quantization (HQ) (Swain and Ballard, 1991) and the

use of different color spaces (Paschos, 2001) in a

whole-image level (Hai et al., 2006) or a multi-scale

framework (Li et al., 2005) for color; Gray level

co-occurrence matrices (GLCM) (Srivastava et al.,

2005), fractal dimension (Chaudhuri and Sarkar,

1995), histograms of oriented gradients (HOG) (Won

et al., 2002) and wavelets (Karkanis et al., 2001) for

texture; and MPEG-7 descriptors and others for shape

(Coimbra and Cunha, 2006) should be adapted to run

online in order to provide a efﬁcient characteriza-

tion of our system. This orientation towards the real

time performance of discriminant features is one of

the most relevant challenges of our current line of re-

search.

Figure 2: Three polyps: ﬂat, peduncular and mixed. Pedun-

cular polyps are prone to their development into malignant

cancers.

In recent works we presented preliminary results

of discriminative features and classiﬁcation systems

for colon cancer detection and its a posteriori char-

acterization in different types of polyps. Figure 2

shows examples of a) peduncular,b) ﬂat and c) mixed

polyps, which have a different degree of clinical rel-

evance, as appearing in our latest contribution (Vi-

lari˜no et al., 2007). Our next step is, hence, to

put these techniques in the horizon of real-time per-

formance. In addition, we argue that the use of

mixed strategies (color/texture/shape) for characteri-

zation would potentially provide a reacher informa-

tion than each of them alone, and the study of the

selection of appropriate features represents an open

ﬁeld of research.

4 POST-INTERVENTIONAL

QUALITY METRICS

We state that computer vision-based techniques can

be used in order to provide post-interventionalmetrics

in an automatic way. We deﬁne four main lines or

research in this area:

1. Automatic estimation of presence of intestinal

content for the assessment of the quality of the

preparation.

2. Percentage of the intestine visualized.

3. Measure of the intervention time.

4. Automatic detection of the different moments

of the intervention (introduction of the endo-

scope/withdrawal).

5. Characterization of the motion of the endoscope.

All these points can be gathered into a general

framework which could be stated in the following

way: The deﬁnition of a 2D map of the patient’s gut

in order to have patient-oriented representation of the

colon, together with their singular features, lesions,

etc. In order to get this wide target, color, textural and

motion features must be used together. Color appears

as a main cue for intestinal content characterization

BIODEVICES 2009 - International Conference on Biomedical Electronics and Devices

322

Figure 3: Eye-tracker conﬁguration system.

(Vilari˜no et al., 2006). Texture can provide a descrip-

tion of the intestinal folds, and motion characteriza-

tion is essential for the measure of transition times. In

a addition, deeper studies must be performed in order

to build up a large database of cases which allows us

to carry out statistical test with the aim of unveiling

the most suitable features and indicators for quality

measures.

5 EYE-TRACING INFORMATION

FOR SKILL ASSESSMENT

The last line of research we introduce in this paper is

related to the assessment of the skills of the trainees

in endoscopy training programs. We hold that using

both the information of the trajectories of the endo-

scopes -which can be potentially obtained by means

of the analysis of the camera movement- and the

tracking of the gaze position -which can be obtained

by an eye-tracker device-, we can provide objective

indicators for the evaluation of the skills. Moreover,

it will be possible to characterize those skills sepa-

rately: ability in endoscope manipulation vs. visual

search.

An eye tracker consists of an adjustable device

with a set of cameras pointing towards the user’s eyes.

The eye tracking procedure involves the recording of

the gaze position and the translation of this values into

the location on the image where the user is focusing

the attention. Figure 3 showsa scheme of the EyeLink

II eye-tracker which was used for our experiments.

The video signal from the cameras is sent to a host

PC in which the eye tracker software is installed, and

which communicates to the experiment PC in which

the visual stimuli are shown. The pupil position is

then translated into gaze position on the stimuli screen

at a very high speed (up to one record each four mil-

liseconds).

6 RESULTS

In this position paper, we would like to highlight pre-

liminary results which were obtained in the colon

cancer characterization and eye-tracker based analy-

sis. The computation of automatic quality assessment

measures is a key part of our ongoing research, and

it will be analyzed in more detail in the Discussion

section.

The eye-tracker information can be used to la-

bel massive data by using the gaze position of the

specialist while screening the colonoscopy video off-

line (in a post-interventional session). Those areas

the specialist is steering the sight towards are asso-

ciated to visual salient features which are of inter-

est for the colonoscopists. This has two main con-

sequences: On the one hand, we can build up an an-

notated database of clinical cases in which the posi-

tion of the observer’s gaze determines the position of

the cancer screened. On the other hand, we can use

the information of the the gaze trajectory to charac-

terize the abilities of the experts in the visual search.

In our case, we built up a database consisting of 6

cases, which is freely accessible for the public and

from the following link:

hhtp://www.cs.tcd.ie/

colon/colon_et_database.zip

We used this database to for the experiments as-

sociated to the colon cancer detection and character-

ization. We applied support vector machine classi-

ﬁers (SVM) (Vapnik, 1995) in order to distinguish

between polyp images and random images of 6 dif-

ferent videos. Each image was 500x700 pixels. We

manually selected those frames were the polyps were

present, and then we trained the classiﬁer with exam-

ples of polyps and non-polyps patches of 128x128

pixels by using the gray-level image, achieving a

working point of around 80% of both sensitivity and

speciﬁcity. The visual analysis of these data pro-

vided promising results in the clustering of the dif-

ferent types of polyps into three basic types (pedun-

cular, ﬂat and mixed). For a further analysis of these

results we refer to the IbPria conference proceedings

(Vilari˜no et al., 2007).

Regarding the analysis of the visual trajectories

obtained from the eye-tracker, we carried out a new

set of experiments in which 20 full colonoscopy stud-

ies were annotated in a two-stage strategy: 1) First,

the expert manually selected those sections of the

video study in which the expert detected the presence

of cancer -this action was performed by clicking and

holding a mouse during the visualization of the polyp-

. 2) Then, only those parts of the video selected by the

specialist in the previous stage were screened with the

eye-tracker. We repeated this experiment with differ-

QUALITY ASSESSMENT IN COLONOSCOPY - New Challenges Through Computer Vision-based Systems

323

ent experts and novicesin order to look for differences

in annotation. The results, which are to be deeply ana-

lyzed in a current study to be submitted in a forthcom-

ing publication, point out that experts and trainees

show different behaviors in the visual search in terms

of reaction time, frequency of saccades, geometry of

the 3D trajectories of the gaze position, etc. These

preliminary results show statistical signiﬁcance, and

they must be contrasted in terms of inter- and intra-

observer variability. This is one of the most relevant

drawbacks of this kind of studies since, the time avail-

ability of the physicians is a major constraint for large

validation studies.

7 DISCUSSION

The preliminary results shown above pave the way

to the computation of quality assessment measures.

For the case of polyps detection, the main aim is to

provide the physician with candidates of polyps to

be analyzed during the intervention. Different statis-

tics, such as the number of polyp-candidates analyzed

and not analyzed, total time consumed in this analy-

sis, etc., can be pulled out in order to obtain objective

indicators. These statistics and indicators must be de-

ﬁned together with the physicians in order to get clin-

ical signiﬁcance and the appropriate tolerance levels.

For the case of eye-tracker data, statistically sig-

niﬁcant indicators of the physicians’ visualization

skills are of relevant importance in a two-fold way:

First, we would be able to provide the physician with

objective metrics that measure high level skills, such

as reaction time, search activity, robustness of the

search pattern, etc., together with general indicators

such as miss rate of polyps. In addition to the for-

mer, we make it possible to decouple the manoeuver-

ing ability from the visual search skills in an objective

way. This information, together with the motion anal-

ysis, can provide indicators regarding the smoothness

of the trajectories that the endoscope performs.

Finally, the automatic estimation of presence of

intestinal content, the quality of the preparation, the

percentage of the intestine visualized, the measure of

the intervention time, and the automatic detection of

the introduction and withdrawal stages must be put

into a clinical framework of quality assessment. This

provides both a study-based assessment and, into a

historical archive, a log of the physician’s indicators

along different interventions. In order to get this done,

a full study is being performed currently together with

the Royal College of Surgeons of Ireland, funded by

an Enterprise Ireland project of the Irish government

for the acquisition and annotation of a full database

of 100 cases of colonoscopy showing colon cancer in

high deﬁnition videos. This database comprises sev-

eral tera bytes of video data and its corresponding an-

notation both in manual and eye-tracked versions by

several specialists with different levels of expertise.

8 CONCLUSIONS

Quality assessment of colonoscopy videos is a rele-

vant issue, since the evaluation of different aspects

of the intervention, by providing objective indicators,

sets up the foundations for a control and reduction of

miss rates in colon cancer detection. We proposed an

approach of quality assessment by means of computer

vision-based techniques which is underpinned by: 1)

The automatic suggestion of potential candidates of

colon cancer during the intervention time, 2) The au-

tomatic computation of objective quality metrics after

the intervention, and 3) The use of eye-tracking in-

formation in order to provide metrics to evaluate the

skills of the physician both in the visual search and

in the endoscope manoeuvering. Preliminary results

showed the suitability of such techniques for polyp

detection and experts vs. trainees discrimination. The

deeper analysis of quality metrics and their correla-

tion is devoted to a further piece of research whose

study our team is carrying out currently.

ACKNOWLEDGEMENTS

We would like to acknowledge the collaboration of

M.D Steven Patchett and M.D Hugh Mulcahy, from

the Royal College of Physicians of Ireland, for their

collaboration in the clinical study of this work. This

research is supported by Enterprise Ireland contracts

PC-2006-038 and CFTD-2006-216.

REFERENCES

American Cancer Society (2007). Cancer Facts and Figure.

American CancerSociety.

Chaudhuri, B. and Sarkar, N. (1995). Texture segmentation

using fractal dimension. IEEE Trans. Pattern Anal.

and Mach. Intel., 17:72–77.

Coimbra, M. and Cunha, J. (2006). MPEG-7 visual descrip-

tors: Contributions for automated feature extraction

in capsule endoscopy. IEEE Transactions on Circuits

ans Systems for Video Technology, 16(5):628–637.

Eddy, D. (1990). Screening for colorectal cancer. Annals of

Internal Medicine, 113:373–384.

BIODEVICES 2009 - International Conference on Biomedical Electronics and Devices

324

Fireman, Z., Glukhovsky, A., et al. (2002). Wireless cap-

sule endoscopy. Israel Medical Association Journal,

4:717719.

Gokturk, S. and Tomasi, C. o. (2001). A statistical 3-d pat-

tern processing method for computer-aided detection

of polyps in ct colonography. IEEE Transactions on

Medical Imaging, 20(12):1251–1260.

Hai, V., Echigo, T., et al. (2006). Adaptive control of video

display for diagnostic assistance by analysis of cap-

sule endoscopic images. In ICPR, volume III, pages

980–983.

Hulka, J. and Reich, H. (1994). Textbook of Laparoscopy.

W.B. Saunders, Phipladelphia, PA, 2 edition.

Hunter, J. and Sackier, J. (1993). Minimally invasive

surgery. McGraw-Hill, New York.

Iddan, G. and Meron, G. e. a. (2000). Wireless capsule

endoscopy. Nature, 405:4–7.

Karkanis, S., Iakovidis, D., et al. (2001). Detection of le-

sions in endoscopic video using textural descriptors on

wavelet domain supported by artiﬁcial neural network

architectures. In ICIP, volume II, pages 833–836.

Kato, H. and Baron, J. (2003). Electronic Videoendoscopy.

Jpapan: Harwood Academic Publisher.

Kudo, S. and Kashida, H. e. a. (2000). Colonoscopic diag-

nosisi and management of nonpolypod and early col-

orectal cancer. World Journal of Surgery, 24(9):1081–

1090.

Li, P., Chan, K., et al. (2005). Learning a multi-size patch-

based hybrid kernel machine ensemble for abnormal

region detection in colonoscopic images. In CVPR,

volume II, pages 670–675.

Liang, J., Higgins, W., et al. (2004). Introduction to the spe-

cial section on virtual endoscopy. IEEE Transactions

on Medical Imaging, 23(11):1333–1334.

Ministry of Health Singapore (1990). MOH Annual Report.

Ministry of Health Singapore.

Nagasako, K., Fijimori, Y., et al. (1998). Atlas of gas-

troenterologic endoscopy by High-Resolution Video-

endoscopy. Igaku-Shoin Ltd. Tokyo.

Nagata, S., Tanaka, K., et al. (2000). Pit pattern diag-

nosis of early colorectal carcinoma by magnifying

colonoscopy: clinical and histological implications.

International Journal of Oncology, 16(5):927–934.

O’Brien, M., Winawer, S., et al. (1990). The national

polyp study: Patient and polyp characteristics asso-

ciated with high-grade dysplasia in colorectoral ade-

nomas. Gastroenterology, 98:371–379.

Pabby, A., Schoen, R., et al. (2005). Analysis of colorec-

tal cancer occurence during surveillance colonoscopy

in the dietary prevention trial. Gastrointestinal en-

doscopy, 61(3):385–391.

Paris Workshop Participants (2003). The Paris endospcopic

classiﬁcation of superﬁcial neoplastic lesions. Gas-

trointestinal Endoscopy (Supplement), 58(6):3–23.

Paschos, G. (2001). Perceptually uniform color spaces for

color texture analysis: empirical evaluation. IEEE

Trans. Image Processing, 10:932–937.

Rembacken, B., Fuji, T., et al. (2000). Flat and de-

pressed colonic neoplasms: A pospective study of

1000 colonoscopies in the uk. Lancet, 17:1157–1167.

Saitoh, Y., Waxman, A., et al. (2001). Prevalence and dis-

tinctive biologic features of ﬂat colorectal adenomas

in a north american population. Gastroenterology,

120:1657–1665.

Srivastava, S., Rodriguez, J., et al. (2005). Analysis of con-

focal microendoscope images for automatic detection

of ovarian cancer. In ICIP, volume I, pages 1113–

1116.

Swain, M. and Ballard, D. (1991). Color indexing. Journal

of computer vision, 7(1):11–32.

Thoeni, R. and Laufer, I. (1994). Textbook of gastrointesti-

nal radiology, chapter Polyps and Cancer, page 1160.

Philadelphia, PA: Sanders.

Tjoa, M. and Krishnan, S. (2003). Feature extraction for the

analysis of colon status from the endoscopic images.

BioMedical Engineering OnLine, 2(9).

U.S. Department of Health and Human Services (2003).

Cdc, the importance of prevention and early detection

cdc. http://www.cdc.gov/cancer/colorectal/.

Vapnik, V. (1995). The nature of statistical learning theory.

Springer.

Vilari˜no, F. (2006). A machine learning approach for in-

testinal motility assessment with capsule endoscopy.

PhD thesis, Dept. Computer Science, Universitat Au-

tonoma de Barcelona, Spain.

Vilari˜no, F., Lacey, G., et al. (2007). Automatic labeling

of colonoscopy video for cancer detection. LNCS,

4477:290–297.

Vilari˜no, F., Spyridonos, P., et al. (2006). Automatic de-

tection of intestinal juices in wireless capsule video

endoscopy. In Pattern Recognition, ICPR, volume 4,

pages 719– 722.

Vilari˜no, F., Spyridonos, P., et al. (2009). Intestinal motil-

ity assessment with video capsule endoscopy: Au-

tomatic annotation of phasic intestinal contractions.

IEEE Trans. on Medical Imaging (in press).

Winawer, S., Fletcher, R., et al. (1997). Colorectal cancer

screening: Clinical guideliness and rationale. Gas-

troenterology, 112:594–642.

Winawer, S., Zauber, A., et al. (1993). Prevention of col-

orectal cancer by colonoscopy polypectomy. New

England Journal of Medicine, 329:977–981.

Won, C., Park, D., et al. (2002). Efﬁcient use of MPEG-7

edge histogram descriptor. ETRI, 24(1):23–30.

Zheng, M., Krishnan, S., et al. (2005). A fusion-based clin-

ical decision support for disease diagnosis from endo-

scopic images. Computers in Biology and Medicine,

35(3):259–274.

QUALITY ASSESSMENT IN COLONOSCOPY - New Challenges Through Computer Vision-based Systems

325