ENDOBRONCHIAL TUMOR MASS INDICATION IN

VIDEOBRONCHOSCOPY

Block based Analysis

Artur Przelaskowski, Rafal Jozwiak

Institute of Radioelectronics, Warsaw University of Technology, Nowowiejska 15/19, 00-665, Warsaw, Poland

Tomasz Zielinski

Department of Telecommunications, AGH University of Science and Technology, Al. Mickiewicza 30, 30-059, Krakow, Poland

Mariusz Duplaga

Collegium Medicum, Jagiellonian University in Krakow, Sw. Anny 12, 31-008, Krakow, Poland

Keywords:

Bronchoscopy, Pattern recognition, Feature selection, Multiscale image processing.

Abstract:

Computer-assisted interpretation of bronchial neoplastic lesion is an innovative but exceptionally challeng-

ing task due to highly diversiﬁed pathology appearance, video quality limitations and the role of subjective

assessment of the endobronchial images. This work is focused on various manifestations of endobronchial

tumors in acquired image sequences, bronchoscope navigation, artifacts, lightening and reﬂections, changing

color dominants and unstable focus conditions. Proposed method of neoplasmatic areas indication was based

on three steps of video analysis: a) informative frame selection, b) block-based unsupervised determining

of enlarged textual activity, c) recognition of potentially tumor tissue, based on feature selection in different

domains of transformed image and Support Vector Machine (SVM) classiﬁcation. Prior to all of these pro-

cedures, wavelet-based image processing was applied to extract texture image for further analysis. Proposed

method was veriﬁed with a reference image dataset containing diversiﬁed endobronchial tumor patterns. Ob-

tained results reveal high accuracy for independent classiﬁcation of individual (single video record) forms of

endobronchial tumor patterns. The overall accuracy for whole dataset of 888 test blocks reached 100%. Less

complex (approximately two times) procedure including initial blocks of interests selection reached accuracy

of 96%.

1 INTRODUCTION

Recent advances in video technology enable for

highly effective and safe diagnostic and therapeutic

procedures of limited invasiveness, which is one of

a key postulates in modern medicine. The quest for

more sophisticated techniques in endoscopy is closely

related to this trend (Duplaga, 2007). Even though,

endoscopic examination remains stressful procedure

for a patient and its outcome depends strongly on

physician’s skills and his subjective assessment of en-

doscopic images. Video recordings of endoscopic

procedures stored in digital libraries are the source of

diversiﬁed and often ambiguous information in terms

of computer-assisted analysis. For example, video

recordings of endoscopic examinations (e.g. gas-

troscopy, colonoscopy, bronchoscopy, etc.) contain

not only images of normal and pathological endolu-

minal structures or diagnostic and therapeutic proce-

dures, but also many poor quality or completely non–

informative frames (e.g. blurred, out-of-focus, dis-

torted, etc.).

1.1 Bronchoscopy

Bronchoﬁberoscopy is one of key diagnostic proce-

dures employed in respiratory medicine and enabling

direct visualization of the endoluminal structure of

tracheobronchial tree. There are many indications

when bronchoﬁberoscopy is performed, but in the

536

Przelaskowski A., Jozwiak R., Zielinski T. and Duplaga M. (2010).

ENDOBRONCHIAL TUMOR MASS INDICATION IN VIDEOBRONCHOSCOPY - Block based Analysis.

In Proceedings of the International Conference on Computer Vision Theory and Applications, pages 536-542

DOI: 10.5220/0002924405360542

 SciTePress

case of lung cancer suspicion it is obligatory. The

procedure is usually accompanied by other diagnos-

tic modalities enabling tissue sampling for pathologic

evaluation. The progress in video technology had also

its impact on bronchoscopy both in terms of available

equipment and the scope of possible diagnostic and

therapeutic techniques. Modern video bronchoﬁbero-

scopes contain an integral video camera system at the

distal end and illumination system based on optical

ﬁbers assuring appropriate visibility of endobronchial

structures. Endobronchial image is captured by the

camera and displayed on the screen which can be con-

veniently placed in the bronchoscopy lab. The ex-

perience of performing physician considerably inﬂu-

ences the effectiveness of the procedure and usually

progress in the skill depends on the intensity of super-

vised training. The assessment of bronchoscopic im-

ages is poorly standardized and currently relies solely

on procedure logbooks and subjective letters of com-

petency (Bowling, 2007).

Video recording of bronchoscopic examinations

demonstrate many common features with natural

video sequences, e.g. general image features and nat-

ural content perception, color space, textural features,

data dynamics, dominant objects properties. Different

parts of bronchoscopic examinations differ in move-

ment characteristics - from slow motion to dynamic

video with fast camera movement across variable di-

agnostic content. Detection of pathological changes

in bronchoscopy frame which comes after a sequence

of normal images may be diagnostically challenging

task.

Video recordings of bronchoscopy procedures are

characterized by a considerable number of illegi-

ble sequences. Furthermore, great part of recorded

frames brings images of normal tracheobronchial tree

which must be also inspected during examination.

While the content of the frames cannot be recognized

and interpreted by any means because of limited ac-

quisition conditioning or case-dependent speciﬁc im-

age appearance, it is impossible to extract from it

any diagnostic information. Such frames with unrec-

ognized content were described as non-informative

frames (Hwang, 2007). The appearance of these

frames is highly diversiﬁed due to many factors in-

ﬂuencing the quality of the endobronchial image. Es-

sentially, we can distinguish: out-of-focusframes (oc-

curred due to wrong camera position - too far/too

close focus into/from mucosa of bronchi), blurred

(motion blur due to rapid endoluminal camera move-

ment), with sanguination (due to pathology presence

or as a result of sampling of suspected tissues for

pathology evaluation) and bubbled (as a result of cam-

era lens cleaning). Informative frames carry poten-

tially important amount of diagnostic information.

However informative frames are not equal in the

sense of visual quality, which change rapidly depend-

ing on current camera situation, position and move-

ment. For example, any camera movement introduces

versatile amount of motion blur (rapid or very fast

camera movement can even lead to total loss of frame

readability), while coverage of the camera lens with

ﬂuids or secretions results in loss of global focus con-

dition. Despite of the diversiﬁed visual quality, most

of informative frames collected in medical video li-

brary represent statistically normal images of differ-

ent anatomical parts of tracheobronchial tree. The

most signiﬁcant video sequences in terms of patho-

logical ﬁndings make relatively small part of reg-

istered frames. The appearance of various patho-

logic manifestations remains strongly highly diversi-

ﬁed both inter-class (between different types of le-

sions) as well as intra-class (between different man-

ifestations of the same type of pathology). Addition-

ally every informative frame can be potentially af-

fected by additional artifacts or distortions (e.g. spec-

ular reﬂections, lightening, etc.) which additionally

hamper process of analysis. Examples of the con-

tents of bronchoscopy video recordings, including

non-informative and informative frames, different le-

sions manifestations and their diversiﬁcation as well

as possible distortions are presented in Fig. 1.

1.2 Pathology Recognition in

Bronchoscopy Video

Computer-based tools developed for bronchoscopy

support were designed with the objective of ad-

vanced visualization aspects like virtual bron-

choscopy (Chung, 2006; Duplaga, 2005) or video

(camera motion) tracking (Rai, 2006; Mori, 2002).

Our work is focused on computer assisted automatic

detection of lesions in bronchoscopy video. To the

best of our knowledge, similar researches have been

reported so far only for endoscopic modalities within

gastrointestinal tract e.g. colonoscopy (Iakovidis,

2006), but not for bronchoscopy.

Among different pathologic endobronchial man-

ifestations, detection of the features of the neoplas-

tic process remains top interest. Endobronchial tumor

mass constitutes relatively common manifestation of

tumors affecting proximal parts of tracheobronchial

tree - Fig. 2. Endobronchial tumor represents it-

self usually as a mass of different consistency in most

cases with dense vasculature and having pink to pur-

plish color.

The purpose of our work was to develop an ef-

ﬁcient method for computer assisted interpretation

ENDOBRONCHIAL TUMOR MASS INDICATION IN VIDEOBRONCHOSCOPY - Block based Analysis

537

Figure 1: The examples of the content of bronchoscopy

video. The ﬁrst row shows examples of non-informative

frames. The second row illustrates informative frames,

readable with good visual quality. Example of onerous vi-

sual quality diversiﬁcation is presented in the third row (ad-

jacent frames from the same part of video examination). Fi-

nally, various bronchial lesions are presented in the fourth

row (from left: tracheal stenosis, extravasations, widened

main carina and mucous ulceration).

Figure 2: Examples of different endobronchial tumors.

Pathology were outlined manually by experienced pul-

monary medicine specialist.

of neoplastic lesions in bronchoscopy images. Sug-

gested method exploits different image preprocess-

ing techniques and is concentrated on texture-based

image analysis. We proposed normalization of ana-

lyzed images, texture feature extraction in different

domains (image, frequency, wavelet and contourlet),

feature selection and unsupervised as well as super-

vised (SVM) classiﬁcation, carried out at different

stages of video content analysis. Implemented algo-

rithms were optimized and veriﬁed experimentally.

2 MATERIAL AND METHODS

General task of diagnostically signiﬁcant pattern

recognition in bronchoscopic images is tough, chal-

lenging and technical conditions dependent because

of mentioned above, seriously limited quality of im-

age information. To succeed, recognition of en-

dobronchial tumor mass was based on an analysis

of bronchoscopy video frames according to precise

successive selection of speciﬁed regions of interests

(ROIs): informative frames, blocks of important con-

tent and blocks of potentially mass appearance.

Because dominant image feature used for recog-

nition was texture, ROIs selection was based gen-

erally on textural data characteristics complimented

with the factors of energy distribution in different data

domains, entropy-based characteristics of stochastic

information and other statistical measures of domi-

nant signal trends.

2.1 Block-based Method Description

Proposed method of neoplasmatic areas indication is

based on three steps of video analysis:

• IFS procedure, i.e. Informative Frame Selection,

based on a whole frame characteristics of possi-

ble artifacts, unusual data and dominant texture

recognition (SVM classiﬁer);

• BUD method, i.e. Block-based Unsupervised De-

termination of enlarged textual activity areas in

the IFS frames, based on energy distribution anal-

ysis and directional image characteristics in polar

2D Fourier space;

• BRT method, i.e. Block-based Recognition of Tu-

mor tissue, based on feature extraction and 48 fea-

ture selection in image, frequency, wavelet, and

contourlet domains, followed by the SVM classi-

ﬁcation.

Prior all of these successive procedures, wavelet-

based image preprocessing was applied to extract tex-

ture image for further analysis.

2.1.1 Texture Extraction Preprocessing

A sequence of the following procedures was used:

• bronchoscopic frame normalization by successive

conversions: a crop of source image area with

mirror fulﬁlling of irregular corners, and conver-

sion to grayscale image (G I) - see Fig. 3;

• contrast-limited adaptive histogram equalization

(CLAHE) of G I with 4×4 pixel blocks and con-

trast function adjusted to 0.02: CL

4×4,0.02

(G I) =

;

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

538

• distorted wavelet synthesis of smoothed source

image I

= W

−1

(W (I

)) and differential, tex-

ture image estimation as: I

= |I

−I

|, where

W is 2-scale dyadic wavelet transformation based

on non-perfect reconstruction ﬁlter bank;

• adaptive histogram equalization of texture image

as follows: I

= CL

4×4,0.02

Exemplary results of texture extraction procedure

were presented in Fig. 3.

Figure 3: The effects of successive image preprocess-

ing procedures: (left-right, top-down) source broncho

frame, cropped image window, image with fulﬁlled corners,

greyscale normalized version (G I), adaptively enhanced

G I, textures extracted by wavelet-based procedure, texture

enhanced by CLAHE.

Histogram equalization is used for imaging

conditioning-invarianceand texture extraction even in

weak signal areas. CLAHE enhances local contrast of

images by transforming the values into the intensity

image. It operates on small data blocks so that the his-

togram of the output region approximately matches

the uniform histogram. The neighboring blocks are

bilineary interpolated to eliminate induced disconti-

nuities on the block boundaries. The contrast, espe-

cially in homogeneous areas, is limited by contrast

enhancement limit parameter to avoid amplifying the

actual image noise or insigniﬁcant textures.

Multiresolution Signal Analysis according to

wavelet-based concept is typically implemented

with speciﬁc types of digital ﬁlter banks (FBs)

known as two- channel perfect reconstruction (PR)

ﬁlter banks. Those ﬁlters are associated with scal-

ing functions (low pass one h) and the wavelets

(high pass one g) of the transform kernel accord-

ing to the following two equations (scaling and

wavelet, respectively): φ(t) =

√

∑

φ(2t −n) and

ψ(t) =

√

∑

φ(2t −n). Conditions of the perfect

reconstructions (Y is almost, i.e. according to the

assumed precision, equal to X) with l delays for

two-channel FB are as follows:

h(z)

h(−z) + g(z) ˜g(−z) = 0 (1a)

h(z)

h(z) + g(z) ˜g(z) = 2z

−l

(1b)

Generally, wavelet decomposition requires the ﬁl-

ters to be FIR (ﬁnite impulse response) and linearly

phased to form orthogonal FBs. However, only Haar

ﬁlter fulﬁll such requirements. Often used solution

is biorthogonal FBs with insigniﬁcant redundancy of

wavelet representation. For texture-oriented image

processing we decided to design orthogonal FB by

softening PR condition. The ﬁrst term (eq. 1a) tra-

ditionally called the alias (cancellation) term is of-

ten fulﬁlled by using quadrature mirror ﬁlters (QMFs)

with conditions: h(z) = ˜g(−z) and g(z) = −

h(−z), as

we did. However, the second term (eq. 1b) called

the distortion elimination term was used to control

the distortion introduced in data processing to ex-

tract basic (lower frequency) signal content. Result-

ing ﬁlter proposition was spline non–PR FB with

h = [1/4,1/2,1/4] and miror g = [−1/4,1/2,−1/4]

for signal smoothing by wavelet preprocessing (Prze-

laskowski, 2007).

2.1.2 Texture Characteristics and Recognition

We considered energy distribution characteristics

across scales and subbands of wavelet domain diver-

siﬁed signiﬁcantly classiﬁed tissue. Different classes

of wavelet energy based features and histogram-based

features from normalized wavelet coefﬁcients were

used. Moreover, entropy features (based on memory-

less and joint source) for subbands compositions and

homogeneity, correlation, energy and contrast of suc-

cessive scale co-occurrence matrix of quantized coef-

ﬁcients were applied. SVM with optimized kernels

and quality criteria was applied for classiﬁcation and

feature reduction procedures.

More precisely, the following textural features

were estimated:

• in image domain:

– statistical (variance, kurtosis, skewness, 0-

order entropy, energy)

– based on co-occurrence matrix (joint entropy,

contrast, correlation, energy, homogeneity)

– Tamura textural features (coarseness, direction-

ality, contrast) (Tamura, 1978)

• in wavelet domain (symlet2 from nearly symmet-

rical wavelets, 2 scales of decomposition):

– energy of approximation related to the energy

of details

– distribution of detail energy and entropy across

scales

ENDOBRONCHIAL TUMOR MASS INDICATION IN VIDEOBRONCHOSCOPY - Block based Analysis

539

– joint entropy of distribution if max magnitude

details for successive scales

– joint entropy, contrast, correlation, energy, ho-

mogeneity for co-occurrence matrix of quan-

tized max magnitude details

– max detail value of different subbands related

to mean approximation energy

• in polar 2D Fourier domain:

– statistical moments of angle histogram of coef-

ﬁcient magnitudes

– energy of angle histogram

– parameters of polynomial approximation of an-

gle histogram

• in contourlet domain – entropy of directional

spectrum of two scales

Selected textural feature spaces were used for IFS

and BRT procedures.

Informative frame selection (IFS) was based on a

whole image I

characteristics for possible artifacts,

regions of unusual data (non–informative) and infor-

mative regions with dominant texture recognition. 48

textural feature space was initially used for frame

classiﬁcation. Informative frames were extracted ac-

cording to supervised classiﬁer of over 800 training

frames. SVM classiﬁer with regularization and radial

kernel was optimized for universal broncho applica-

tions.

Block-based unsupervised determination (BUD) of

enlarged textual activity areas was used for fast deﬁn-

ing of block of interests. Idea of second step of tumor

extraction was high risk area segmentation through

block-oriented cover of the image. Blocks of 50×50

pixels were veriﬁed basing on energy distribution

analysis and directional image characteristics in po-

lar 2D Fourier space. Two phase threshold selection:

for energy distribution factors and directional factors

was optimized for over 1000 test cases.

Block-based recognition of tumor tissue (BRT) was

designed for ﬁnal recognition of selected, active im-

age blocks as potentially covered by tumor mass tis-

sue. Recognition scheme was optimized for large test

set of several thousand of test BUD blocks basing on

textual feature extraction and selection for effective

case classiﬁcation. SVM procedure with regulariza-

tion and radial kernel was used to classify each BUD

block as diagnostically suspected of having pathology

symptoms or normal.

Graphical form of selection and recognition re-

sults was used for indication of frame status as non–

informative (purple mark in top left corner) or infor-

mative (lack of purple mark), while informativeframe

regions could be covered by red (i.e. BUD) or yellow

(i.e. BRT) blocks.

The proposed block-based method assisting intra-

bronchial tumor mass recognition was implemented

in MATLAB environment. For classiﬁcation purpose

we used standard SVM classiﬁer from Matlab Bioin-

formatics Toolbox. Additionally we used some ex-

tra procedures form Contourlet Toolbox (Do, 2005),

Beamlab (Donoho, 2001) and Wavelab (Buckheit,

1995).

2.2 Test Bronchoscopic Material

The proposed method was veriﬁed with a reference

image data set containing different bronchoscopy

video content exempliﬁcation (e.g. varied patholo-

gies, diversiﬁed anatomical bronchi structures, ther-

apeutic and diagnostic interventions, possible arti-

facts, etc.) clinically selected by experienced pul-

monary medicine specialist from near 600 recorded

and stored videos of bronchoscopy examinations. A

reference set of close to 1300 images containing

informative/non-informativeframes was used to asses

IFS efﬁciency.

For the purpose of BUD and BRT assessment 14

diversiﬁed cases of intrabronchial tumor manifesta-

tions were considered, each containing several espe-

cially selected frames with additional manual pathol-

ogy outline. Every frame was divided into blocks

of size 50x50, where each block was categorized to

norm or pathology according to available physician

outline. Near 900 selected blocks were ﬁnally used

as a test set. Among these blocks, a small set of 38

blocks of one distinct case of intrabronchial tumor

was selected. 16 of these blocks contained representa-

tive patterns of clear, high quality manifestation of the

tumor completed with 22 blocks of surrounding, nor-

mal tissues. Such test sets were used for BUD/BRT

veriﬁcation and optimization.

3 EXPERIMENTS AND RESULTS

Examples of results achieved for the Matlab imple-

mentation of the described algorithm were presented

and explained in Fig. 4. Illustration of the method ad-

justing to concreteexamination and user requirements

were outlined in Fig. 5.

IFS step tested against 1228 test frames gave even

100% accuracy of informative/non-informative selec-

tion with optimally adjusted radial kernel of SVM for

complete space of extracted features. Optimization

procedure indicated that IFS efﬁciency moderately

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

540

Figure 4: Exemplary results of the proposed block-based

method applied to selected video examination with en-

dobronchial tumor mass (reference frame with outlined

pathology is presented in top left corner). Other frames of

ﬁrst row, marked with purple block, were classiﬁed as non-

informative (ﬁrst stage of algorithm). Red blocks represent

areas with high texture information activity (related directly

to second stage of proposed method) while yellow blocks

marked out presence of pathology area (related directly to

third stage of proposed method). Second row illustrates in-

formative frames without clearly visible pathology (frames

preceding appearance of pathology). Yellow blocks notice-

able on third frame in this row (from left) represent false

positive (FP) indications. Frames with visible pathology

are present in third row. As we can see, pathology exis-

tence are indicated more or less precisely, but mainly cor-

rectly (yellow block are present in places corresponding to

area outlined originally on reference frame). Some miss-

classiﬁcation examples can be seen on ﬁrst frame (form left)

- two yellow blocks located above pathology area.

Table 1: IFS effectiveness for selected sets of textural fea-

tures and adjusted classiﬁer; six the most useful features

are: entropy of contourlet coefﬁcients, ratio of approxima-

tion to detail wavelet coefﬁcients, Tamura coarseness, en-

tropy of wavelet details in successive scales, entropy and

variance of extracted textures in image domain.

Number of

the features Classiﬁer Sensitivity Accuracy

48 SVM/rbf 1 1

48 SVM/linear .9 .92

24 SVM/rbf 1 1

24 SVM/linear .89 .92

6 SVM/rbf .91 .94

6 SVM/linear 0.87 0.87

depends on classiﬁcation procedure and even signiﬁ-

cant reduction of the number of textural features. Ex-

emplar results were presented in Tab. 1.

Sensitivity of BUD method was optimized in re-

lation to high enough speciﬁcity through adjusting of

Figure 5: Impact of method parametrization for overall de-

tection efﬁciency. Each row illustrates results for different

set of method parameters. Depending on the selected set of

initial presets proposed method is more or less sensitive. Se-

lection of this parameters should set a compromise between

method sensitivity and speciﬁcity.

Table 2: Adjusting of BUD parameters to select the best

balance between sensitivity and speciﬁcity of the method.

Threshold values of 4.35/1 (4.35 for normalized energy dis-

tribution and 1 for directional image characteristics) were

used ﬁnally.

Two threshold

values of BUD Sensitivity Speciﬁcity

4.35/4 .65 .58

4/4 .70 .36

4.35/3 .81 .44

4.35/1 .89 .38

4/2 .95 .1

two selective thresholds. Diversiﬁed in local energy

test set of 888 image blocks was used to make the

procedure more universal, sufﬁciently efﬁcient as on-

line detector of suspicious to tumor areas and selec-

tive enough for more precise supervising of BRT pro-

cedure.

Satisfying sensitivity of 94% with speciﬁcity up

to 86% was achieved for more distinct test set of 38

selected blocks. The balance range between sensitiv-

ity and speciﬁcity for total test set of 888 blocks was

presented in Tab. 2. Maximum accuracy was limited

to 61%.

Accuracy of BRT method was adjusted to 98% for

888 test control or pathology blocks. Selected 30 tex-

tural feature space was used. Six the most useful fea-

tures are: entropy of contourlet coefﬁcients, mean of

angle histogram and the parameter of polynomial ap-

proximation of angle histogram in polar 2D Fourier

domain, mean local entropy in image-texture domain

ENDOBRONCHIAL TUMOR MASS INDICATION IN VIDEOBRONCHOSCOPY - Block based Analysis

541

combined with texture energy, global entropy and en-

ergy in image-texture domain. However, accuracy for

these 6 features was only 75%. Accuracy of BRT for

more ﬂexible and universal linear kernel of SVM was

only 76%.

Obtained results reveal high accuracy for indepen-

dent classiﬁcation of individual differential forms of

endobronchial tumor patterns, especially basing on

time consuming IFS-BRT procedure. The overall ac-

curacy for whole dataset of 888 test blocks reached

100%. Thus, time consuming IFS-BRT combination

that assumed feature extraction and SVM classiﬁca-

tion for each frame and next successively for all frame

blocks is effective enough to ﬁt classiﬁcation rules

to tumor detection problem. Less complex (approx-

imately ﬁve times) procedure of complete IFS-BUD-

BRT reached accuracy of 96%.

All veriﬁed procedures were designed to analyze

bronchoscopic video in order to indicate the blocks of

high susceptibility to tumor mass. A way of frame

and frame block selection to be analyzed depends

on application requirements. Because of computa-

tional complexity of BRT, which is fundamental pro-

cedure for tumor recognition, ad-hoc section method

of blocks of interests, similar to BUD or other inter-

active methods based on human-computer interfaces,

are useful for close to on-line application.

4 CONCLUSIONS

Clinical usefulness of the proposed method should

be further tested in conditions of bronchoscopy suit.

Reliable experimental procedure strongly depends on

signiﬁcantly diversiﬁed technical conditions of bron-

choﬁberoscopes and test cases. Moreover, the fea-

sibility of this method may be affected by the lim-

ited standardization of the procedure and signiﬁcant

role of subjective assessment. However, automatic

indications in almost on-line mode (second stage of

the method) or more reliable in off-line mode (full

method application) seems to be useful as an assistant

tool for more careful bronchoscopic video analysis.

Objectiﬁed indicators of special regions of interests

are useful for standardized protocol design, compara-

tive analysis and education of inexperienced doctors.

As far as the authors are aware, it is the ﬁrst attempt of

development of the tool based on the automatic prob-

able pathology indication supporting bronchoscopic

examination.

REFERENCES

Duplaga, M., Leszczuk, M., Przelaskowski, A., Janowski,

L. and Zieliski, T. (2007). Bronchovid - zin-

tegrowany system wspomagajcy diagnostyk bron-

choskopow. Przegld Lekarski 64:42-48.

Bowling, M., Downie, G., Wahidi, M. and Conforti, J.

(2007). Self-Assessment Of Bronchoscopic Skills In

First Year Pulmonary Fellows. Chest Vol. 132, Issue

Hwang, S., Oh, J., Lee, J., Tavanapong, W., de Groen, P. C.

and Wong, J. (2007). Informative Frame Classiﬁcation

for Endoscopy Video. Medical Image Analysis Vol.

11, No 2:100-127.

Chung, A. J., Deligianni, F., Shah, P., Wells, A. and Yang,

G. Z. (2006). Patient Speciﬁc Bronchoscopy Visu-

alisation through BRDF Estimation and Disocclusion

Correction. IEEE Transactions of Medical Imaging

25(4):503- 513.

Duplaga, M. and Socha, M. (2005). Aplikacja oparta na bib-

liotece VTK wspomagajca zabiegi bronchoskopowe.

Bio-Algorithms and Med-Systems I(l/2):191-196.

Rai, L., Merritt, S. A. and Higgins, W. E. (2006). Real-

time image-based guidance method for lung-cancer

assessment. IEEE Conf. Computer Vision and Pattern

Recognition 2:2437-2444.

Mori, K., Deguchi, D., Sugiyama, J., Suenaga, Y., Toriwaki,

J., Maurer, C. R. Jr, Takabatake, H. and Natori, H.

(2005). Tracking of a bronchoscope using epipolar

geometry analysis and intensity-based image registra-

tion of real and virtual endoscopic images. Med. Im-

age Anal. 6:321-365.

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

(2006). An Intelligent System for Automatic De-

tection of Gastrointestinal Adenomas in Video En-

doscopy Computers in Biology and Medicine. Vol. 36,

10:1084-1103.

Przelaskowski, A., Bargiel, P., Sklinda K. and Zwierzynska

E. (2007). Ischemic stroke modeling: multiscale ex-

traction of hypodense signs Lecture Notes in Artiﬁcial

Intelligence 4482:171-181, Springer Verlag.

Tamura, H., Mori, S. and Yamawaki, T. (1978). Textu-

ral features corresponding to visual perception IEEE

Trans. Systems, Man. and Cybern. Vol. 8, 6:460-472 .

Do, M. N. and Vetterli, M. (2005). The contourlet trans-

form: an efﬁcient directional multiresolution image

representation IEEE Trans Image Proces. Vol. 14,

12:2091-2106 .

Donoho, D. L. and Huo, X. (2001). Beamlets and Multi-

scale Image Analysis Computational Science and En-

gineering, Multiscale and Multiresolution Methods,

Springer.

Buckheit, J. B. and Donoho, D. L. (2005). WaveLab and

Reproducible Research Dept. of Statistics, Stanford

University, Tech. Rep. 474.

VISAPP 2010 - International Conference on Computer Vision Theory and Applications

542