Professional Documents
Culture Documents
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2399502, IEEE Journal of Biomedical and Health Informatics
2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2399502, IEEE Journal of Biomedical and Health Informatics
In this paper, we propose words based color histogram for experimental results validating the proposed intelligent system
bleeding detection in WCE images. Our method is an extension are presented and discussed in Section V. Finally, we draw
of Bag of Words method [12, 13]. In order to make the most of some conclusions and discussions in Section VI.
the color information of the bleeding images, we calculate the
color words by applying K-means clustering procedure to the II. METHOD OVERVIEW
pixel represented WCE images in the specific color space. Then The flowchart of the proposed method is depicted in Fig. 1
each WCE image is characterized as histogram of the cluster and it comprises two major steps. In the first step of bleeding
centers (named words based color histogram) to represent the frame classification, we propose words based color histograms
feature vector. Finally, support vector machine (SVM) [14] and to represent WCE images. Then we focus on the bleeding area
K nearest neighbor (KNN) [15] are utilized as classifiers to localization. We calculate the first stage saliency map based on
detect bleeding frames. In the second stage, we focus on the observation of the bleeding areas in the different color
localization of the bleeding areas in the bleeding frames. Since spaces and the second stage saliency map based on the natural
the components of various color spaces possess different property of bleeding areas in the RGB color space. Finally we
information, we inspect the bleeding images under different localize the bleeding areas through the fusion strategy of the
color spaces like RGB, HSI/HSV, CMYK, CIELAB, YUV, saliency maps and an appropriate threshold.
XYZ [16] and select the components that highlight the bleeding
areas. Then we create the first stage saliency map by combing III. BLEEDING FRAME DETECTION
these components together to strengthen the suspicious regions.
In addition, natural saliency lies in the visual contrast [17, 18]. A. ROI Extraction
Thus we extract the second stage saliency map from the prior sWCE images are often obscured by the large black
that if the color information of the region shows large similarity background and obvious bounders as shown in Fig. 2 (a), (b),
to the red color, then this region should possess high saliency thus the image features extracted from the entire image will
value. Finally applying an appropriate fusion strategy of these reflect the visual contamination presented in the image. To
two conspicuity maps and automatic threshold, we are able to address this factor, we outline the maximum square inscribed in
localize the bleeding area. the circular image as region of interest (ROI) without loss of the
Our main contributions can be summarized in the following major image information. The size of the obtained ROI is
two aspects. 180180 from the original image with the size of 256256. It
can be found that the extracted ROIs are satisfactory since they
1) We propose the words based color histogram to represent
demonstrate the major image features and provide a good
the images. That is, we make use of middle level features rather
characterization and description of the WCE images. The ROI
than low level features that are generally used in the literatures images replace the original ones for the following processing,
[7, 9-11]. Moreover, we study carefully the influence of the making the feature extraction procedure much easier.
color spaces, cluster centers and different classification
methods in terms of the classification performance.
2) We propose a two-stage saliency extraction method to
localize the bleeding areas in WCE images. Since these
two-stage saliency maps highlight the bleeding regions and
separate bleeding mucosa from the uninformative parts, we
could obtain the bleeding area candidates successfully.
The remainder of this paper is organized as follows. Section
II outlines the proposed method. Section III introduces the
Fig. 2 Illustration of ROI extraction.
bleeding frame classification method while Section IV
discusses the localization of the bleeding areas. The
2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2399502, IEEE Journal of Biomedical and Health Informatics
2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2399502, IEEE Journal of Biomedical and Health Informatics
2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2399502, IEEE Journal of Biomedical and Health Informatics
V. EXPERIMENT RESULTS
A. Image acquisition and experimental setup
In our experiments, we conducted experiments on 2400
WCE images that consist of 400 bleeding frames and 2000
normal frames. These images were extracted from 10 different
patients videos and manually annotated by gastroenterologists.
The effective resolution of these images is 256*256.
B. Experiment Results for Bleeding Frames Detection
1) Parameters Selection
The first experiment was designed to evaluate the
discriminative power of the proposed feature and choose the
best parameters in the classification of the bleeding WCE
images. Fig. 6. Feature analysis. (a) Normal images and the corresponding feature
The most important two parameters in the construction of the histogram. (b) Bleeding images and the corresponding feature histogram.
word based color histogram are the selection of the vocabulary
size K and the color space. To evaluate how these two 3) Comparison to the other methods
parameters influence the classification performance, we To further evaluate the performance of the proposed new
conduct the coupled experiments. We tune K from 10 to 100 feature, we compared it with the six state-of-the-art color
with 10 increments in the experiment and the results for the features. The first two comparison methods directly used the
different classification methods SVM and KNN in different HSV histogram and YCbCr histogram to describe images.
color spaces are showed in the Fig. 5. To be mentioned here, we These histograms are produced by dividing the colors in the
image into specific number of bins and counting the number of
apply a 10-fold cross validation method to validate our
image pixels in each bin. For the HSV histogram, it quantizes
proposed method [27].
the whole color space into 72 main colors (dividing H into eight
We first analyzed the classification performance with
parts, S and V into three parts) while for the YCbCr histogram,
different classification algorithms. It can be found that the 125 bins are used to represent image features (dividing Y, Cb,
SVM method (Fig. 5 (a)-(c)) is much more suitable than the Cr into 5 parts). The method in [10] and [11] has already been
KNN method (Fig. 5 (d)-(f)) for bleeding classification in the mentioned in the introduction part and these two papers are the
WCE images because the SVM method shows relative higher latest related papers focused on the bleeding detection. We also
accuracy, specificity and sensitivity in all of the color spaces. compared the proposed method with latest feature coding
We then studied the classification performance with the SVM method: sparse coding [28] and Locality-constrained linear
classifier. As illustrated in Fig. 5 (a)-(c), there are some coding (LLC) [29]. In these two methods, after obtaining the
differences among the four color spaces. Overall the YCbCr cluster centers, instead of represent image by vector
color space exhibits the best classification performance while quantization, they encode each point in the image by the sparse
the HSV color space is not suitable for bleeding detection with coding and LLC method to obtain the final histogram. We
lowest accuracy compared with other color spaces. implemented these algorithms on our datasets and the
In addition, the cluster center size does affect the corresponding comparison results are showed in Table I.
classification performance. In the YcbCr color space, the TABLE I
classification performance increases first and then decreases PERFORMANCE COMPARISON OF DIFFERENT FEATURES (%)
when the cluster size increases. The best classification result is HSV YCbCr [10] [11] [28] [29] Our
histo- histo- me-
achieved with cluster centers 80 and the corresponding gram gram thod
accuracy and sensitivity reach 95.75%, 92.00%. The AUC of Acc. 92.37 86.25 90.00 83.00 91.63 76.63 95.75
the classification with the best parameter setting is 0.9771, Sen. 84.75 91.50 81.75 78.00 94.75 91.25 92.00
Spec. 93.90 85.20 91.65 84.00 91.00 62.00 96.50
validating the proposed feature has a good capability to classify
Time(s) 75.93 422.27 35.76 530.43 4171.5 28737 293.43
bleeding images.
2) Feature analysis It can be found in Table I, the proposed method shows better
To show the effectiveness of our features, we plotted two performance to those of the HSV histogram method with an
examples of the feature histogram with cluster number 80 for improvement 3.38% and 7.25% of in accuracy and sensitivity,
the normal images and bleeding images in the Fig. 6. The respectively, while to those of YCbCr histogram with an
cluster centers 1-40 are obtained by the bleeding images while improvement 9.5% and 0.5%. This result demonstrates that the
the cluster centers 41-80 are obtained from the normal images. words based color histogram is superior compared with the
We could find that the bleeding image and the normal image traditional color histograms to classify the bleeding frames. In
addition, our method gets higher accuracy, sensitivity and
show significant different words based color histograms.
specificity than the existed method [10] and [11] in the bleeding
frame classification task. Compared with the latest feature
coding method [28] and [29], our method not only shows better
accuracy and sensitivity, but also is much faster than sparse
2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2399502, IEEE Journal of Biomedical and Health Informatics
Fig. 5 Classification performance with different color spaces and classification method. (a)-(c) the classification performance with SVM. (d)-(f) shows the
classification performance with KNN.
coding and LLC, which demonstrating our method is more 2) Qualitative Analysis
effective in clinical situation. We fixed w1 0.8 to illustrate our proposed saliency
C. Experiment Results for Bleeding Localization methods visually by the first example image given in Fig. 5. Fig.
1) Quantitative Analysis 9(a) is the saliency map generated through means of channel
After obtaining the bleeding frames, we focused on the mixer while Fig. 9(b) is obtained by prior knowledge that the
localization of the bleeding areas in this part. Our model region with red color should assign large saliency value. It is
calculates two stage saliency maps and fuses them together to obvious that these two saliency maps provide good ability to
detect the bleeding areas. Evaluation of the fusion strategy is an identify bleeding regions. After the fusion strategy, the final
important step in the assessment of the model performance. We saliency map is shown in Fig. 9(c). In this figure, the bleeding
mucosa appears brighter than normal mucosa region, hence it
tested w1 in the Eq. 6 for the values {0,0.2,0.4, 0.6,0.8,1} to
can be easily extracted by an appropriate threshold. Fig. 9(d) is
obtain the final saliency maps and then applied the Otsus [30] the binary bleeding mask image obtained through the Otsus
threshold method on the calculated saliency maps to obtain the threshold procedure.
bleeding areas.
TABLE II
PERFORMANCE COMPARISON OF DIFFERENT FOR w1 BLEEDING
LOCALIZATION (%)
W1 Precision FPR FNR W1 Precision FPR FNR
2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/JBHI.2015.2399502, IEEE Journal of Biomedical and Health Informatics
2168-2194 (c) 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See
http://www.ieee.org/publications_standards/publications/rights/index.html for more information.