Computer-aided classification of colorectal polyps based on vascular patterns: a pilot study

• Introduction
• Methods
• Patients and study design
• Image evaluation by investigator
• Computer-based image evaluation
• Statistics
• Results
• Image evaluation by investigator
• Computer-based image evaluation
• Comparison of results
• Discussion
• Acknowledgments
• References

Background and study aims: Recent studies have shown that narrow-band imaging (NBI) is a powerful diagnostic tool for differentiating between neoplastic and nonneoplastic colorectal polyps. The aim of the present study was to develop and evaluate a computer-based method for automated classification of colorectal polyps on the basis of vascularization features.

Patients and methods: In a prospective pilot study with 128 patients who were undergoing zoom NBI colonoscopy, 209 detected polyps were visualized and subsequently removed for histological analysis. The proposed computer-based method consists of image preprocessing, vessel segmentation, feature extraction, and classification. The results of the automated classification were compared to those of human observers blinded to the histological gold standard.

Results: Consensus decision between the human observers resulted in a sensitivity of 93.8 % and a specificity of 85.7 %. A “safe” decision, i. e., classifying polyps as neoplastic in cases when there was interobserver discrepancy, yielded a sensitivity of 96.9 % and a specificity of 71.4 %. The overall correct classification rates were 91.9 % for the consensus decision and 90.9 % for the safe decision. With ideal settings the computer-based approach achieved a sensitivity of approximately 90 % and a specificity of approximately 70 %, while the overall correct classification rate was 85.3 %. The computer-based classification showed a specificity of 61.2 % when a sensitivity of 93.8 % was selected, and a 53.1 % specificity with a sensitivity of 96.9 %.

Conclusions: Automated classification of colonic polyps on the basis of NBI vascularization features is feasible, but classification by observers is still superior. Further research is needed to clarify whether the performance of the automated classification system can be improved.

Introduction

Based on the concept of the adenoma-carcinoma sequence, most colorectal cancers arise from pre-existing adenomas [1]. Early detection and removal of colorectal adenomas may greatly reduce the incidence of colorectal carcinoma, hence the importance of preventive screening colonoscopies [2]. However, a significant proportion of colorectal polyps are nonneoplastic, without malignant potential, and in such cases polypectomy is usually unnecessary.

Given the complication rate of polypectomy, which is up to 2.7 % of all cases, a noninvasive way of distinguishing between neoplastic and nonneoplastic colorectal lesions would be very useful [3]. Magnifying chromoendoscopy allows neoplastic colon polyps to be distinguished from nonneoplastic ones using the pit-pattern classification by Kudo and colleagues, which has a diagnostic accuracy of about 90 %. However, chromoendoscopy is labor-intensive and the results are very investigator-dependent [4].

Recently, a new technique called narrow-band imaging (NBI) has been developed, which enhances the epithelial microvascular pattern [5]. In a previously published study, we demonstrated that, using NBI illumination with magnification and applying a vascular pattern classification, human observers were able to correctly identify 93.7 % of neoplastic polyps and 89.2 % of nonneoplastic colorectal lesions [6]. These results are comparable to the findings of other groups and confirm the important role of neoangiogenesis in tumor growth [7] [8] [9] [10].

However, classification of colorectal lesions on the basis of vascular patterns is still not objectively standardized, and intra- and interobserver variance is a problem for polyp classification under NBI illumination [11] [12]. Intraobserver variance may be caused by increasing expertise, external distraction, personal feelings and well-being, or stress. Interobserver variance mainly arises from differing levels of expertise. We developed a computer-based vascularization analysis algorithm for colorectal polyps with a subsequent automatic classification step in order to minimize the investigator dependency of classification results, and evaluated the results in a pilot study.

Methods

Patients and study design

The pilot study was designed to detect a difference of 10 % in accuracy with a power of 80 % by a two-sided McNemar test at the significance level α of 5 %, assuming 25 % rate of discordant pairs. Thus, 207 polyps had to be entered into the study. Based on the experience of our previous study [6] we calculated that we would need approximately 5 months to accrue more than 207 colorectal polyps. A total of 641 patients underwent colonoscopy during a period of 5 months. Patients with chronic inflammatory bowel disease, adenomatosis coli, coagulopathy, insufficient bowel preparation, or who had undergone colonoscopy within the last 3 years (except for patients referred for polypectomy of known polyps) were excluded from the study. All patients who did not meet the exclusion criteria were included in the study. Overall, 223 of 641 patients were included in the study and underwent colonoscopy with a NBI zoom endoscope, which can magnify the image to a maximum of 150-fold (CF-Q160ZI, Olympus Medical Systems Europe). In 128 of the 223 patients, a total of 209 polyps were detected, imaged at a magnification of × 100, and removed for histological analysis, which was taken as the gold standard for classification.

The study protocol was approved by the local ethics committee of Aachen University Hospital, RWTH Aachen University, and patients gave their written informed consent to participation in the study.

Image evaluation by investigator

One magnified NBI image of each of the 209 polyps was given to two investigators, who were blinded to histological data, for independent evaluation in a first round. Both observers were asked to classify the vascular pattern according to the intensity and shape of small blood vessels (nonneoplastic: fine capillary pattern, normal size and distribution; neoplastic: increased number, tortuous, corkscrew-type). In cases where there was a discrepancy between the decisions of the two observers, the observers were asked to discuss the image and reach a joint decision (“consensus decision”). In addition, a separate, “safe” decision was made, where the polyp was declared to be an adenoma if either of the two observers came to this conclusion. Thus, only polyps marked by both observers as hyperplasias were denoted as such.

Computer-based image evaluation

The computer-based algorithm was applied to the same images as were given to the two observers. In addition, the computer was supplied with a map on which the polyp surface on the input image was marked. Details of the preprocessing applied, vessel segmentation, and classification of the endoscopic images have been published in recent technical reports [13] [14]. A sequence of images illustrating the algorithm is shown in [Fig. 1].

Briefly, an image taken during colonoscopy is taken ([Fig. 1 a]). The specular reflections are detected and automatically excluded from further analysis ([Fig. 1 b]) [15]. After this step, the image is converted to gray scale: only the green channel is used as it provides the best contrast between polyp surface and blood vessels [16]. Since illumination varies greatly in the colon, the final step of preprocessing is illumination normalization ([Fig. 1 c]). After preprocessing, candidate blood vessels are detected and enhanced using an analysis based on phase symmetry ([Fig. 1 d]) [17]. From the candidate vessels, a region growing algorithm, the fast marching algorithm, is started to expand to the whole vessel lumen [18]. This final segmentation result can be seen in [Fig. 1 e]. The seed point detection and vessel segmentation are quantitatively evaluated for parametrization purposes by comparing their results with blood vessels manually segmented in five polyp images.

From these (intermediate and result) images, several values – called features – characterizing neoplastic and nonneoplastic lesions can be computed. The features mean vessel length, mean vessel circumference, and mean brightness value at image positions with detected blood vessels are then used in the final step – the classification itself. In this case, the term “mean” stands for a value normalized to the lesion’s surface area. These features achieved good results in preliminary tests. The three resulting numbers are concatenated in a (feature) vector of the form (x, y, z) which can be interpreted as a position (coordinates) in a three-dimensional space (“feature space”).

Each new polyp, represented by its feature vector, is assigned to one of the polyp groups (“neoplastic” and “nonneoplastic”). We investigated several algorithms called classifiers that can infer class affiliation on the basis of feature vector position in a three-dimensional feature space, such as linear classifiers, k-nearest-neighbor (k-NN), and support vector machines (SVMs). SVMs [19] [20] [21] are able to separate highly complex distributions flexibly by nonlinear decision rules and perform well on our experimental setup. The classifier training is easy with today’s mathematical resources. The classification process itself is reduced to a few scalar products and comparisons, and therefore the classification is very fast, which is a requirement for a clinical system. Linear classifiers are not able to model complex distributions, while for k-NN classifiers each polyp to be classified has to be compared to all training samples and the computational load therefore rises in a linear fashion with the number of polyps.

In our study the classification was carried out using a leaving-one-out approach. A comparison of possible approaches can be found in Martens and Dardenne [22]. An SVM was trained to discriminate nonneoplastic from neoplastic polyps using features from n – 1 = 208 training samples; the remaining sample was then classified by the SVN. This was done for every image from the 209-image database. The result is the average recognition rate for all 209 samples.

During an intervention, the steps needed to classify one image are to calculate the features for the recently taken image and apply the results to the SVM classifier. Both steps are comparably fast and the current pilot system, which is not optimized for speed and is implemented in MATLAB, takes less than 30 seconds to perform these tasks. Optimization for clinical trials would significantly reduce the calculation time and allow comfortable use during interventions.

Statistics

Sensitivity, specificity, and accuracy were compared by applying the two-sided McNemar test. True-positive lesions were polyps that were correctly classified as neoplastic. A P value below 0.05 was considered statistically significant. Interobserver agreement in distinguishing between neoplastic and nonneoplastic colorectal polyps was determined by calculation of κ values.

Results

The mean size of the 209 polyps was 8.1 mm (range 2 – 40 mm); 163 (78.0 %) of the colorectal lesions detected were polypoid and 46 (22.0 %) were flat lesions. Overall, 49 colorectal lesions (23.44 %) were found to be nonneoplastic upon histological evaluation, whereas 160 (76.56 %) polyps were neoplastic. Of the 49 nonneoplastic polyps, 45 were hyperplastic, 2 inflammatory, and 2 were lymphoid follicles. Among the 160 neoplastic lesions, we found 113 tubular adenomas with low-grade dysplasia, 37 tubulovillous adenomas with low-grade dysplasia, 1 villous adenoma with low-grade dysplasia, 1 tubular adenoma with high-grade dysplasia, 5 tubulovillous adenomas with high-grade dysplasia, and 3 colorectal carcinomas. No serrated adenomas were detected. Following surgery 2 of the 3 colorectal carcinomas were histologically staged as pT1N0M0G2, while the third was staged as T2N1M0G2.

Image evaluation by investigator

Digital images of the 209 polyps were classified according to their vascular patterns. For 9 colorectal lesions there was a discrepancy in vascular pattern analysis between the two investigators, leading to an interobserver agreement with κ values of 0.808. Image evaluation by investigator A resulted in a sensitivity of 95.0 % and a specificity of 79.6 % for differentiation of nonneoplastic and neoplastic lesions. Investigator B achieved a sensitivity of 95.6 % and a specificity of 81.6 %. The consensus decision resulted in an increased specificity of 85.7 % (95 %CI: 72.1 – 93.6), but also in a drop in sensitivity to 93.8 % (95 %CI: 88.5–96.8) ([Table 1]).

Making the “safe” decision in the case of a discrepancy yielded a sensitivity of 96.9 % (95 %CI: 92.5 – 98.8) and a specificity of 71.4 % (95 %CI: 56.5 – 83.0) ([Table 2]).

In the subgroup of 44 polyps that were either high-grade adenoma or (tubulo-)villous lesions, the consensus decision resulted in a sensitivity of 90.9 % and the safe decision in a sensitivity of 95.5 %, whereas in the subgroup of 113 tubular adenomas with low-grade dysplasia, 94.7 % and 97.3 % were correctly classified by consensus decision and safe decision, respectively. There was no significant difference between the diagnostic values of these subgroups.

Computer-based image evaluation

Seed points for the initialization of the subsequent vessel segmentation were detected by phase symmetry analysis [17]. The region growing results gave error rates of 3.5 % and 2.6 % on the nonneoplastic and neoplastic colorectal polyps, respectively, in our segmentation test set.

Applying our algorithm to the realistic test set, 209 colorectal polyps were analyzed. [Fig. 2] shows results for feature 1 (mean vessel length), feature 2 (mean vessel circumference), feature 3 (mean brightness value at image positions with detected blood vessels), and human observers.

Based on a variation of the segmentation and classification parameters, a receiver operating characteristic (ROC) curve was generated. The ROC curve clearly indicates that combining the features increased the classification performance. The combined result clearly shows the largest area under the curve (AUC). Among the 209 colorectal polyps, for example, the computer-based analysis with the combined features achieved a sensitivity of 90 % and a specificity of 70.2 %.

For the subgroup of 44 polyps including cases of either high-grade adenoma or (tubulo-)villous lesions, computer-based classification resulted in a sensitivity of 97.7 %, whereas among 113 tubular adenomas with low-grade dysplasia, 92.0 % were correctly classified by computer-based algorithm. Again, there was no significant difference between the diagnostic values of the two subgroups.

Comparison of results

To directly compare the human and automatic results, we selected the parameters of the algorithm such that the computer-based results exhibited sensitivities identical to those of consensus (93.8 %) and safe decision (96.9 %). For the consensus decision the specificity is 85.7 %. The respective individual specificities are 61.2 % (95 %CI: 46.2 – 74.5) (combined features), 61.2 % (feature 1), 8.2 % (feature 2), and 6.1 % (feature 3) ([Table 1]). For the safe decision the specificity is 71.4 %. The respective individual specificities are 53.1 % (95 %CI: 38.4 – 67.2) (combined features), 36.7 % (feature 1), 4.1 % (feature 2), and 2.0 % (feature 3) ([Table 2]).

There was no significant difference regarding the sensitivities of the different methods between the two subgroups of colorectal lesions with lower and higher risk of cancer development.

Discussion

Detection of neoplastic colorectal polyps is important because of their well-known relationship to cancer development [1]. Removing neoplastic polyps reduces the risk of developing colorectal cancer [2]. However, a significant proportion of the colorectal lesions removed are histologically nonneoplastic and therefore without malignant potential. In these cases polypectomy, which is associated with a risk of bleeding or even colon perforation, would in most cases not have been necessary. In addition to these risks, removal of nonneoplastic lesions not only increases the duration of the colonoscopic procedure, decreasing efficiency, but also results in unnecessary costs. Ideally, the indication for colonoscopic polypectomy should be limited to neoplastic polyps, which have the potential of developing into infiltrating carcinomas. Therefore, a noninvasive way of differentiating between neoplastic and nonneoplastic colorectal lesions would be highly beneficial.

In this prospective study, we were able to confirm, for the first time to the best of our knowledge, that NBI with magnification permits a computer-based evaluation of colorectal polyps based on the vascular patterns with a diagnostic accuracy acceptable for a pilot study. Out of the dataset of 209 polyps containing 49 (23.4 %) nonneoplastic and 160 (76.6 %) neoplastic polyps, the human investigators correctly identified 93.8 % of the neoplastic and 85.7 % of the nonneoplastic lesions. When the two investigators came to different conclusions, a consensus decision resulted in higher specificity but noticeably lower sensitivity. In clinical practice, however, the rate of misclassification of neoplastic lesions would be minimized. Classifying as neoplastic all lesions for which the two reviewers had discrepant results (“safe decision”), a sensitivity of 96.9 % was reached, but simultaneously specificity declined to 71.4 %. The overall correct classification rates for the human investigators were 91.9 % for the consensus decision and 90.9 % for the safe decision.

These results are similar to those of our previous study and matched the findings of other groups, in all of which sensitivities and specificities of approximately 90 % were achieved [6] [7] [8] [9]. In our study the interobserver agreement was high, with κ = 0.808, comparable to the results of other studies [11] [12].

In addition to the interobserver variability due to the subjective nature of current microvessel classification systems, a distinct learning curve in the analysis of vascular pattern has been observed [11]. In the course of their study, Rogart and colleagues demonstrated an improvement of diagnostic accuracy from 74 % to 87 % in distinguishing between neoplastic and nonneoplastic polyps using NBI without magnification [11]. Computer-aided diagnosis using advanced image analysis algorithms could remove some of the subjectivity and allow nonexperts to achieve high levels of diagnostic accuracy.

For this reason, we developed an entirely computer-based system for the analysis of vascular patterns. To this end we identified the following three features that allow differentiation between neoplastic and nonneoplastic colorectal polyps: (i) the mean vessel length, (ii) the mean vessel circumference, and (iii) the mean brightness value at image positions with detected blood vessels. The combination of all three criteria led to the best overall diagnostic accuracy. The computer-based system allows adaptation of the sensitivity and specificity to the operator’s needs. The sensitivity and specificity are, as always, inversely related and an increase in sensitivity causes a decrease in specificity. Among the 209 colorectal polyps, for example, the computer-based analysis achieved a sensitivity and specificity of 90 % and 70.2 %, respectively. This corresponds to an overall correct classification rate of 85.3 %. This diagnostic value indicates that automatic and hence investigator-independent classification of colon polyps is feasible.

In direct comparison with expert-investigator classification, however, computer-aided analysis is still significantly inferior in terms of specificity. With in each case a sensitivity of 93.8 %, the investigators correctly identified 85.7 % of nonneoplastic lesions, whereas the computer-based system achieved a specificity of 61.2 %. One possible explanation for this difference is that the human observers unconsciously took account of additional characteristics other than vascular patterns. In a previous study, we demonstrated that combinations of vascular and mucosal pattern classification using the NBI system resulted in an increased sensitivity for differentiating between neoplastic and nonneoplastic colorectal polyps [6]. The absolute correct classification rates of approximately 91 % for the human observers and 85 % for the computer-based system re-emphasize the feasibility of the computer-based classification approach.

Note that the correct classification rate is higher for the important group of colorectal lesions at increased risk of cancer development in the case of the automated classification system. However, this result is not significant, which might be due to the small sample size in the subgroup.

Future work may involve including other physiological properties, e. g., pit-pattern analysis. Medically indicated characteristics of the vascular pattern, including shape, alignment, and course of blood vessels, may be integrated into the computer-based analysis. In this pilot study we focused on the feasibility of a computer-based classification system and did not aim to maximize the performance by means of specialized features. The computer-based system should also be tested on a database of images containing multiple images of the same polyps to confirm robustness against variations in the image data. Automated location of polyps is currently under development [25]. However, as we were conducting a pilot study on classification performance, we did not apply the segmentation algorithm in this study.

In conclusion, we have presented a method for automatic and hence investigator-independent classification of colorectal polyps based on the patterns of blood vessels. We have applied preprocessing and vessel segmentation to colonoscopic NBI images. We have calculated features based on characteristics of blood vessels and classified the polyps according to these features. The results show that automated classification of colon polyps using NBI vascularization features is feasible but still inferior to classification by human experts. At the present time the system is not suitable for routine clinical use.

Future research will determine the possibility of further improvements and a closing of the described gap in performance. Ideas for future enhancements include further investigations into more features, and combinations of blood vessel features and statistical features [13] [14]. We are also aiming to enhance the imaging quality and the statistical evaluation. Our future work will also include automatic location and segmentation of polyp surfaces in NBI images.

Acknowledgments

This research project was supported by the Excellence Initiative of the German Federal and State Governments and the START Program of the Faculty of Medicine, RWTH Aachen University.

Competing interests: None

References

• 1 Vogelstein B, Fearon E R, Hamilton S R. et al . Genetic alterations during colorectal-tumor development.  N Engl J Med. 1988;  319 525-532
  CrossrefPubMedSearch in Google Scholar
• 2 Winawer S J, Zauber A G, Ho M N. et al . Prevention of colorectal cancer by colonoscopic polypectomy. The National Polyp Study Workgroup.  N Engl J Med. 1993;  329 1977-1981
  CrossrefPubMedSearch in Google Scholar
• 3 Heldwein W, Dollhopf M, Rösch T. et al . The Munich Polypectomy Study (MUPS): prospective analysis of complications and risk factors in 4000 colonic snare polypectomies.  Endoscopy. 2005;  37 1116-1122
  Thieme ConnectPubMedSearch in Google Scholar
• 4 Kudo S, Tamura S, Nakajima T. et al . Diagnosis of colorectal tumorous lesions by magnifying endoscopy.  Gastrointest Endosc. 1996;  44 8-14
  CrossrefPubMedSearch in Google Scholar
• 5 Gono K, Obi T, Yamaguchi M. et al . Appearance of enhanced tissue features in narrow-band endoscopic imaging.  J Biomed Opt. 2004;  9 568-577
  CrossrefPubMedSearch in Google Scholar
• 6 Tischendorf J JW, Wasmuth H E, Koch A. et al . Value of magnifying chromoendoscopy and narrow band imaging (NBI) in classifying colorectal polyps: a prospective controlled study.  Endoscopy. 2007;  39 1092-1096
  Thieme ConnectPubMedSearch in Google Scholar
• 7 Su M Y, Hsu C M, Ho Y P. et al . Comparative study of conventional colonoscopy, chromoendoscopy, and narrow-band imaging systems in differential diagnosis of neoplastic and nonneoplastic colonic polyps.  Am J Gastroenterol. 2006;  101 2711-2716
  CrossrefPubMedSearch in Google Scholar
• 8 Chiu H M, Chang C Y, Chen C C. et al . A prospective comparative study of narrow-band imaging, chromoendoscopy, and conventional colonoscopy in the diagnosis of colorectal neoplasia.  Gut. 2007;  56 373-379
  CrossrefPubMedSearch in Google Scholar
• 9 Rastogi A, Bansal A, Wani S. et al . Narrow-band imaging colonoscopy – a pilot feasibility study for the detection of polyps and correlation of surface patterns with polyp histological diagnosis.  Gastrointest Endosc. 2008;  67 280-286
  CrossrefPubMedSearch in Google Scholar
• 10 Risau W. Mechanisms of angiogenesis.  Nature. 1997;  386 671-674
  CrossrefPubMedSearch in Google Scholar
• 11 Rogart J N, Jain D, Siddiqui U D. et al . Narrow-band imaging without high magnification to differentiate polyps during real-time colonoscopy: improvement with experience.  Gastrointest Endosc. 2008;  68 1136-1145
  CrossrefPubMedSearch in Google Scholar
• 12 Sikka S, Ringold D A, Jonnalagadda S. et al . Comparison of white light and narrow band high definition images in predicting colon polyp histology, using standard colonoscopes without optical magnification.  Endoscopy. 2008;  40 818-822
  Thieme ConnectPubMedSearch in Google Scholar
• 13 Gross S, Stehle T, Behrens A. et al . A comparison of blood vessel features and local binary patterns for colorectal polyp classification.  Proc SPIE Medical Imaging. 2009;  72602Q 1-8
  PubMedSearch in Google Scholar
• 14 Stehle T, Auer R, Gross S. et al . Classification of colon polyps in NBI endoscopy using vascularization features.  Proc SPIE Medical Imaging. 2009;  72602S 1-12
  PubMedSearch in Google Scholar
• 15 Stehle T. Removal of specular reflections in endoscopic images.  Polytechnica J Adv Eng. 2006;  46 32-36
  PubMedSearch in Google Scholar
• 16 Hajer J, Kamel H, Noureddine E. Blood vessels segmentation in retina image using mathematical morphology and the STFT analysis.  Inf Commun Technol. 2006;  2 1130-1134
  PubMedSearch in Google Scholar
• 17 Kovesi P. Image features from phase congruency.  Videre J Comput Vis Res. 1999;  1(3) 1-27
  PubMedSearch in Google Scholar
• 18 Sethian J A. A fast marching level set method for monotonically advancing fronts.  Proc Natl Acad Sci USA. 1996;  93 1591-1595
  CrossrefPubMedSearch in Google Scholar
• 19 Cortes C, Vapnik V. Support vector networks.  Machine Learning. 1995;  20(3) 273-297
  PubMedSearch in Google Scholar
• 20 Vapnik V. Statistical learning theory. New York; Wiley 1998
  PubMedSearch in Google Scholar
• 21 Christianini N, Shawe-Taylor J. An introduction to support vector machines and other kernel-based methods. Cambridge; Cambridge University Press 2000
  PubMedSearch in Google Scholar
• 22 Martens H A, Dardenne P. Validation and verification of regression in small data sets.  Chemometrics Intelligent Lab Syst. 1998;  44 99-121
  CrossrefPubMedSearch in Google Scholar
• 23 Atkin W S, Morson B C, Cuzick J. Long-term risk of colorectal cancer after excision of rectosigmoid adenomas.  N Engl J Med. 1992;  326 652-662
  PubMedSearch in Google Scholar
• 24 Lieberman D A, Weiss D G, Harford W V. et al . Five-year colon surveillance after screening colonoscopy.  Gastroenterology. 2007;  33 1077-1085
  PubMedSearch in Google Scholar
• 25 Gross S, Kennel M, Stehle T. et al . Polyp segmentation in NBI colonoscopy.  Bildverarbeitung für die Medizin. 2009;  2009 252-256
  PubMedSearch in Google Scholar

J. J. W. TischendorfMD 

Medical Department III (Gastroenterology, Hepatology and Metabolic Diseases)
University Hospital Aachen
RWTH Aachen University

Pauwelsstr. 30
52074 Aachen
Germany

Fax: +49-241-80-82455

Email: jtischendorf@ukaachen.de

References

• 1 Vogelstein B, Fearon E R, Hamilton S R. et al . Genetic alterations during colorectal-tumor development.  N Engl J Med. 1988;  319 525-532
  CrossrefPubMedSearch in Google Scholar
• 2 Winawer S J, Zauber A G, Ho M N. et al . Prevention of colorectal cancer by colonoscopic polypectomy. The National Polyp Study Workgroup.  N Engl J Med. 1993;  329 1977-1981
  CrossrefPubMedSearch in Google Scholar
• 3 Heldwein W, Dollhopf M, Rösch T. et al . The Munich Polypectomy Study (MUPS): prospective analysis of complications and risk factors in 4000 colonic snare polypectomies.  Endoscopy. 2005;  37 1116-1122
  Thieme ConnectPubMedSearch in Google Scholar
• 4 Kudo S, Tamura S, Nakajima T. et al . Diagnosis of colorectal tumorous lesions by magnifying endoscopy.  Gastrointest Endosc. 1996;  44 8-14
  CrossrefPubMedSearch in Google Scholar
• 5 Gono K, Obi T, Yamaguchi M. et al . Appearance of enhanced tissue features in narrow-band endoscopic imaging.  J Biomed Opt. 2004;  9 568-577
  CrossrefPubMedSearch in Google Scholar
• 6 Tischendorf J JW, Wasmuth H E, Koch A. et al . Value of magnifying chromoendoscopy and narrow band imaging (NBI) in classifying colorectal polyps: a prospective controlled study.  Endoscopy. 2007;  39 1092-1096
  Thieme ConnectPubMedSearch in Google Scholar
• 7 Su M Y, Hsu C M, Ho Y P. et al . Comparative study of conventional colonoscopy, chromoendoscopy, and narrow-band imaging systems in differential diagnosis of neoplastic and nonneoplastic colonic polyps.  Am J Gastroenterol. 2006;  101 2711-2716
  CrossrefPubMedSearch in Google Scholar
• 8 Chiu H M, Chang C Y, Chen C C. et al . A prospective comparative study of narrow-band imaging, chromoendoscopy, and conventional colonoscopy in the diagnosis of colorectal neoplasia.  Gut. 2007;  56 373-379
  CrossrefPubMedSearch in Google Scholar
• 9 Rastogi A, Bansal A, Wani S. et al . Narrow-band imaging colonoscopy – a pilot feasibility study for the detection of polyps and correlation of surface patterns with polyp histological diagnosis.  Gastrointest Endosc. 2008;  67 280-286
  CrossrefPubMedSearch in Google Scholar
• 10 Risau W. Mechanisms of angiogenesis.  Nature. 1997;  386 671-674
  CrossrefPubMedSearch in Google Scholar
• 11 Rogart J N, Jain D, Siddiqui U D. et al . Narrow-band imaging without high magnification to differentiate polyps during real-time colonoscopy: improvement with experience.  Gastrointest Endosc. 2008;  68 1136-1145
  CrossrefPubMedSearch in Google Scholar
• 12 Sikka S, Ringold D A, Jonnalagadda S. et al . Comparison of white light and narrow band high definition images in predicting colon polyp histology, using standard colonoscopes without optical magnification.  Endoscopy. 2008;  40 818-822
  Thieme ConnectPubMedSearch in Google Scholar
• 13 Gross S, Stehle T, Behrens A. et al . A comparison of blood vessel features and local binary patterns for colorectal polyp classification.  Proc SPIE Medical Imaging. 2009;  72602Q 1-8
  PubMedSearch in Google Scholar
• 14 Stehle T, Auer R, Gross S. et al . Classification of colon polyps in NBI endoscopy using vascularization features.  Proc SPIE Medical Imaging. 2009;  72602S 1-12
  PubMedSearch in Google Scholar
• 15 Stehle T. Removal of specular reflections in endoscopic images.  Polytechnica J Adv Eng. 2006;  46 32-36
  PubMedSearch in Google Scholar
• 16 Hajer J, Kamel H, Noureddine E. Blood vessels segmentation in retina image using mathematical morphology and the STFT analysis.  Inf Commun Technol. 2006;  2 1130-1134
  PubMedSearch in Google Scholar
• 17 Kovesi P. Image features from phase congruency.  Videre J Comput Vis Res. 1999;  1(3) 1-27
  PubMedSearch in Google Scholar
• 18 Sethian J A. A fast marching level set method for monotonically advancing fronts.  Proc Natl Acad Sci USA. 1996;  93 1591-1595
  CrossrefPubMedSearch in Google Scholar
• 19 Cortes C, Vapnik V. Support vector networks.  Machine Learning. 1995;  20(3) 273-297
  PubMedSearch in Google Scholar
• 20 Vapnik V. Statistical learning theory. New York; Wiley 1998
  PubMedSearch in Google Scholar
• 21 Christianini N, Shawe-Taylor J. An introduction to support vector machines and other kernel-based methods. Cambridge; Cambridge University Press 2000
  PubMedSearch in Google Scholar
• 22 Martens H A, Dardenne P. Validation and verification of regression in small data sets.  Chemometrics Intelligent Lab Syst. 1998;  44 99-121
  CrossrefPubMedSearch in Google Scholar
• 23 Atkin W S, Morson B C, Cuzick J. Long-term risk of colorectal cancer after excision of rectosigmoid adenomas.  N Engl J Med. 1992;  326 652-662
  PubMedSearch in Google Scholar
• 24 Lieberman D A, Weiss D G, Harford W V. et al . Five-year colon surveillance after screening colonoscopy.  Gastroenterology. 2007;  33 1077-1085
  PubMedSearch in Google Scholar
• 25 Gross S, Kennel M, Stehle T. et al . Polyp segmentation in NBI colonoscopy.  Bildverarbeitung für die Medizin. 2009;  2009 252-256
  PubMedSearch in Google Scholar

J. J. W. TischendorfMD 

Medical Department III (Gastroenterology, Hepatology and Metabolic Diseases)
University Hospital Aachen
RWTH Aachen University

Pauwelsstr. 30
52074 Aachen
Germany

Fax: +49-241-80-82455

Email: jtischendorf@ukaachen.de