1.

Introduction

A video endoscope system is now considered an indispensable tool for the diagnosis of early-stage gastrointestinal disease. In recent years, endoscopic therapy has attracted a great deal of interest because of the advantages of minimally invasive techniques. To improve the chances for total tumor resection by endoscopic means, early detection is mandatory. To detect these early cancers, it is important to precisely observe the color variation and fine structural pattern of the mucosa. Kudo et al.;^{1 2 3} reported that estimating the pathology of a colon polyp is possible by observing the crypt orifice (the so-called pit pattern) of colonic glands with a magnified endoscope. Tobita⁴ reported the usefulness of the classification of minute gastric mucosal patterns, which was proposed by Sakaki,⁵ for diagnosing the depressed type of early gastric cancers.

Chromoscopy has been used widely to precisely observe the fine structural patterns of the mucosa.^{6 7} Chromoscopy is a contrast staining method using a biocompatible dye agent, such as indigo carmine or methylene blue. In the mucosa with glands, the dye agents accumulate within crypt orifices. Although chromoscopy is effective in many applications, this method still has some problems, such as difficulty in achieving complete and even coating of the mucosal surface with the dye, the extra cost of the equipment for dye spraying, the dye agent itself, and the extra time required to perform the procedure. Kumagai et al.;⁸ reported that observation of an intrapapillary capillary loop pattern is potentially useful for diagnosing the depth of invasion of early esophageal cancer. Several investigations have been conducted on the diagnosis of early gastric cancer through observation of the capillary patterns on the gastric mucosa.⁹ However, chromoscopy is not useful for enhancing the imaging of the capillary patterns. To resolve the problems of chromoscopy, we proposed narrow-band imaging (NBI).¹⁰ This is a medical video endoscope system utilizing narrow-band illumination. Narrow-band illuminations are created with three optical interference filters in which the bandwidths are spectrally narrowed.

Sano et al.;¹¹ conducted preliminary clinical tests of NBI during colonoscopy. They reported that imaging of the colonic mucosa with NBI provides enhanced pit pattern contrast compared with ordinary observation. However, a quantitative understanding of the relation between spectral features of the illumination and tissue features of the observational image is still lacking. Kienle et al.;¹² investigated how the wavelength of the illumination affects the reproduced color of the vein under the skin using a tissue phantom. They used a thin glass tube phantom containing whole blood in a lipid colloid. It is, however, unclear how the contrast of the thinner vasculature, such as the capillaries, depends on the center wavelength of the filtered illumination.

2.

Experiment

This study was performed to determine the basis for tissue imaging by narrow-band illumination and to confirm the clinical benefits of NBI. Three experiments were conducted as follows.

3.

Materials and Methods

3.1.

Medical Video Endoscope System

The medical video endoscope system shown in Fig. 1 was used in the present study. The light source unit (CLV-U40, Olympus Optical, Tokyo, Japan) has a xenon lamp and a filter disk. Three optical interference filters are mounted on the filter disk. Red, green, blue (R,G,B) filters are ordinarily used and divided visible wavelength ranges into three bands. The video endoscope has a light guide of an optical fiber bundle and a monochrome CCD at its tip. R,G,B illuminations irradiate the mucosa sequentially via the light guide as the filter disk rotates. The monochromatic image signal corresponding to each filtered illumination is formed by synchronizing the sequential illumination process and the conversion process by the CCD. Analog signals read from the CCD are transformed sequentially to digital signals in the video processor (CV240, Olympus Optical). To display a color image on a monitor (OEV-142, Olympus Optical), each monochromatic image signal (band image) is assigned to an R/G/B color channel. Then, the R/G/B color signals are sent out to the monitor as National Television System Committee (NTSC) television signals. These signals become input signals of the frame grabber (METOR II/MC4; Matrox Electronic Systems, Dorval, Canada). The frame grabber can digitize television signals with a sampling of 640×480 pixels. Moreover, the gray levels in each pixel are quantized with integer values from 0 to 255.

Figure 1

Medical video endoscope system using the RGB sequential illumination method.

3.2.

Experiment 1: Contrast Between the Vascular Pattern and the Adjacent Mucosa of the Tongue

The contrast between the vascular pattern and the adjacent mucosa of the underside of the human tongue was evaluated quantitatively. The structure of the vascular pattern of the human tongue is similar to that of the gastrointestinal mucosa,^{14 15} therefore, this tissue was appropriate for our study. Evaluations were performed using eight different illuminations. Five narrow-band filters and three broadband filters were used. The center wavelengths and bandwidths are listed in Table 1. F6, F7, and F8 have the same specifications as the filters utilized in the ordinary light source unit. F1, F4, and F5 were selected to investigate the influence of bandwidth by comparison with F6, F7, and F8. Narrow-band filters F2 and F3 were selected for evaluation of the changes in contrast within a short wavelength range. In this range, the molar excitation coefficient of the hemoglobin changes most sharply in the visible range.¹³

Table 1

Center wavelength and bandwidth of filters.
	F1	F2	F3	F4	F5	F6	F7	F8
Wavelength	415 nm	445 nm	500 nm	540 nm	600 nm	420 nm	540 nm	610 nm
Bandwidth	30 nm	30 nm	30 nm	20 nm	20 nm	100 nm	80 nm	80 nm

A magnified video endoscope (GIF-Q240Z, Olympus was used to observe the vascular pattern. This apparatus has an 80-fold optical magnifying capacity. Operating the knob at the handle grip allows switching between the wide view and the telescopic view. Experiments observing the human tongue were performed with the telescopic view to visualize the capillaries. In the telescopic view, because of the small depth of field of the optics, it is difficult to fix the focus. Measuring the contrast requires sharp focus. A transparent silicone cap was attached to the tip of the video endoscope. A stable sharp focus was obtained by placing the tip of the cap gently on the tissue. Although using the cap changes the blood flow slightly, we assumed that the change in blood flow has little effect on the contrast of the vascular pattern. Indeed, placing the cap gently on the mucosa was accompanied by little if any change in the vascular pattern. Fixing the gain of the video processor maintained the signal-to-noise ratio of images at a uniform level. The dark noise was automatically subtracted from the analog signal of the CCD. The images were stored in a personal computer as digital data. Before measuring the contrast, the tone of the image was transformed since its gamma equals unity.

Figure 2 shows the schema of the vascular patterns. A dense capillary bed is typically found in superficial layers of the mucosa, and thicker blood vessels run in deeper layers. The difference in blood vessel depth may influence the contrast of the image. Therefore, to examine the influence of the depth of blood vessels on contrast, we classified the vascular patterns into three types by thickness and depth in the mucosa. The capillaries found in the superficial mucosa were classified as class 1 (C1 pattern). The thicker vasculature in deeper layers was classified as class 2 (C2 pattern). Thick veins were classified as class 3 (C3 pattern). The thicknesses of C1, C2 and C3 patterns are about 10 to 20, 20 to 50, and 200 to 500 μm, respectively.

Figure 2

Structure of the blood vessels in the human tongue. C1 indicates the fine capillaries in the superficial mucosa. C2 indicates the thick capillaries in regions deeper than C1. C3 indicates the thick veins in regions deeper than C1 and C2.

In this study, the contrast was defined as the luminance ratio according to Weber’s law,¹⁶ as shown in Eq. (1). In this definition, a pixel value of 8 bits (0 to 255) was used to calculate substitution for the luminance.

Eq. (1)

$$ctr=|g_v-g_m|/g_m,$$

where g denotes the pixel value of the recorded image and subscripts v and m refer to the vascular pattern and its adjacent mucosa, respectively. The positions of blood vessels were selected manually by observation. After the position of the blood vessel was selected, the position of its adjacent mucosa was selected from an area where no blood vessel could be seen. Two hundred and fifty such selections were conducted for the three classes of vascular pattern. As a result, 250 contrast values were obtained in each filter for the class of vascular pattern. The differences in average contrast values among the different filters and the different classes of blood vessel were evaluated statistically. In our experiments, it could not be assumed that variances of the pair to be compared were equal. Therefore, the test statistic of Aspin-Welch [Eq. (2)] was employed¹⁷:

Eq. (2)

$$t_o=\frac{{\tilde{c}}_1-{\tilde{c}}_2}{\sqrt{\frac{s_1^2}{n_1}+\frac{s_2^2}{n_2}}},$$

where c˜₁ and c˜₂ denote averages calculated from 250 contrast values, s₁ ² and s₂ ² denote variances, and n₁ and n₂ denote number of samples (n₁=n₂=250). This is distributed approximately as a Student’s t distribution with the degree of freedom v shown as Eq. (3).

Eq. (3)

$$\upsilon =\frac{{\left(\frac{s_1^2}{n_1}+\frac{s_2^2}{n_2}\right)}^2}{\frac{{\left(\frac{s_1^2}{n_1}\right)}^2}{n_1-1}+\frac{{\left(\frac{s_2^2}{n_2}\right)}^2}{n_2-1}}.$$

In this study, one-sided hypothesis tests were performed. The p value was deduced from the t distribution table as the probability p=Pr{|t|⩾t_o}/2. A p value of 1 or less (p<0.01) was taken to indicate significance.

4.

Results

4.1.

Experiment 1: Contrast Between the Vascular Pattern and the Adjacent Mucosa of the Tongue

The images produced by the five narrow-band filters and three broadband filters are shown in Fig. 3. The relationships between the vascular patterns and classification are also illustrated schematically in Fig. 3. Considering images F1 to F5, it is clear that the appearance of the vascular patterns depends on the class of the blood vessels and the center wavelength of the narrow-band filter. C1 patterns were seen in the F1 image. However, fine features, such as the tip of the capillary, were blurred in the F2 image. Furthermore, such fine features were difficult to see in the F4 image and could not be seen at all in the F5 image. In F3 and F4 images, although the C1 pattern was blurred, the C2 pattern could be seen more clearly than in the F1 image. In the F5 image, the C1 and C2 patterns were blurred. The C3 pattern was also blurred in the F5 image. However, the F5 image seemed to be appropriate for visualizing the C3 pattern. Examination of F1 and F6 indicated that narrowing the bandwidth improves visualization of the C1 pattern. The effects of narrowing bandwidth on visualization of the C2 and C3 patterns are unclear.

Figure 3

Narrow-band images and broadband images of the underside of the human tongue. (Top) Narrow-band images (F1 to F5). (Bottom) Broadband images (F6 to F8). (Bottom left) Diagrammatic representation of the blood vessel classes.

Figure 4 shows the average values of the contrast in each class and in each band image. The asterisks in the figure indicate that the contrast could not be calculated because the vascular patterns could not be determined manually. The error bars represent 1 standard deviation. The average values were in good agreement with the results of the observations described earlier. The p values are listed in Table 4.

Figure 4

Contrast between each class of blood vessels and the neighboring mucosa. The asterisks in the figure indicate that the contrast could not be calculated because the blood vessel could not be seen. The error bars indicate 1 standard deviation n=250.

Table 4

p values of statistical analyses.
	F1	F2	F3	F4	F5
C1:F1	—	2.73×10−27	4.52×10−110	1.40×10−40	9.71×10−139
C2:F4	4.47×10−20	8.49×10−20	0.36	—	7.70×10−58
C3:F5	*	*	9.21×10−26	1.29×10−5	—
* Contrast could not be calculated.

The results for the C1 patterns demonstrated a statistically significant difference in a comparison of F1 images, which have a higher contrast than the other band images (p<0.01). The results for the C2 pattern indicated no statistically significant differences between F4 and F3 images (p=0.36). The results also demonstrated a statistically significant difference with F4 images, which have a higher contrast than the other band images (p<0.01). The results for C3 patterns demonstrated a statistically significant difference in F5 images, which have a higher contrast of C3 vasculature than the other band images (p<0.01). For the C1 pattern, F1 images showed significantly higher contrast for blood vessels than F6 images (p=8.52×10⁻¹¹). For the C2 pattern, F4 images had a higher contrast for blood vessels than F7 images, and this difference was statistically significant (p=9.30×10⁻¹⁰). For the C3 pattern, F5 images showed a significantly higher contrast for blood vessels than F8 images (p=1.63×10⁻⁹⁷).

4.2.

Experiment 2: Appearance of C1 Pattern in Colored Images of the Tongue

Figure 5 shows colored images of the underside mucosa of the human tongue. The color in Figs. 5(b) and 5(c) is a natural reddish tissue color. On the other hand, that shown in Fig. 5(a) is different from the others. However, the C1 pattern could be seen more clearly in Fig. 5(a). The upper row in Table 5 shows the results of CIE 1976 L ^* a ^* b ^* color difference (dE) between the C1 pattern and the adjacent mucosa. The value of dE in Set B was significantly larger than both that of Set A (p=2.77×10⁻²¹) and that of Set C (p=6.18×10⁻³²). The lower row in Table 5 shows the results of L ^* ctr. The value of L ^* ctr in Set A was significantly larger than those of Set B (p=2.28×10⁻²⁸) and Set C (p=1.83×10⁻²⁵).

Figure 5

Color images of the underside of the human tongue. (a) Filter Set A. (b) Filter Set B. (c) Conventional broadband filters.

Table 5

CIE 1976 L * a * b * color difference and contrast of L * of C1 pattern.
	Set A	Set B	Set C
dE	13.96±3.98	18.19±5.47	12.17±5.24
L * ctr	0.19±0.07	0.12±0.06	0.12±0.06

The results from CIE 1976 L ^* a ^* b ^* color differences indicated that Set B is an appropriate filter set for color imaging. On the other hand, the results from L ^* ctr indicated that Set A is an appropriate filter set for color imaging. The color images shown in Fig. 5 indicate that L ^* ctr is more suitable than CIE 1976 L ^* a ^* b ^* color differences for evaluating the appearance of the blood vessels. This is discussed in detail in Sec. 5.

4.3.

Experiment 3: Clinical Tests

4.3.1.

Colonoscopy

As an example of a clinical test, images obtained in a case of familial adenomatosis coli are shown in Fig. 6. In the NBI image, the typical branching pattern of blood vessels can be seen as a blue pattern. Moreover, small polyps of familial adenomatosis coli can be seen more clearly than those in the conventional image. The small polyps in NBI images can be seen as brown staining. The pathology of most small polyps are adenomatous, and such polyps show angiogenesis in the superficial layer of the mucosa, which can be detected in the F1 and F2 images. As a result, most of the small polyps can be seen clearly as brown staining. Table 6 shows the results from scoring by a gastroenterologist using a visual scale. The results indicated that NBI can more clearly define tumor margins than conventional illumination, and this difference is statistically significant (p=3.557×10⁻⁴).

Figure 6

(a) Conventional image and (b) NBI image of familial adenomatosis coli.

Table 6

Scores of visualization of colonic small polyps.
	Unevaluated Score 1	Unclear Score 2	Clear Score 3	Total Score	Average Score	p value
Conventional	6	16	5	26	0.963	3.557×10−4
NBI	2	7	18	43	1.593

4.3.2.

Barrett’s esophageal endoscopy

As an example of a clinical test, Images obtained in a case of BE are shown in Fig. 7. The images are telescopic views obtained using the magnified function of the video endoscope. The region labeled “A” is the normal esophagus mucosa. The region marked “B” is the BE mucosa and shows a textured pattern. It is clear that NBI allows visualization of the pit pattern more clearly than conventional illumination. Table 7 shows the results of scoring by a gastroenterologist using a visual scale. The results agreed with comparative observations of images between conventional illumination and NBI. NBI showed a significant advantage in visualization of the pit pattern of the BE mucosa compared with conventional illumination (p=2.24×10⁻⁴).

Figure 7

(a) Conventional image and (b) NBI image of Barrett’s esophagus.

Table 7

Scores of visualization of Barrett’s pit pattern.
	Completely Impossible Score 1	Visualization Just possible Score 2	Fairly Sharp Visualization Score 3	Complete Visualization Possible Score 4	Total Score	Average Score	p value
Conventional	0	4	17	0	55	2.620	2.246×10−4
NBI	0	0	13	8	71	3.381

5.

Discussion

In narrow-band images, the contrast of the C1 pattern shows a maximum at the maximum absorption wavelength of hemoglobin (415 nm). In addition, the contrast of the C1 pattern was blurred in the wavelength range outside that of hemoglobin absorption. On the other hand, the contrast of C2 and C3 patterns showed maxima in wavelength ranges different from the maximum absorption wavelength of hemoglobin. This is due to the differences in the position within the mucosa of the different classes of vascular pattern. The C1 pattern is in the superficial layer and therefore changes in its contrast with wavelength reflect the absorption spectrum of hemoglobin. On the other hand, C2 and C3 patterns are in the middle or deep layers of the mucosa, and therefore scattering and absorption by tissue and capillaries in the upper layer rather than in the middle or deep layers markedly affect the contrast. To allow visualization of the C2 or C3 pattern, the illumination has to penetrate the upper layer and reach the middle or deep layer. C2 and C3 patterns must have the thickness to compensate for the slight hemoglobin absorption. Therefore a vascular pattern imaged in the longer wavelength range is thicker than the C1 pattern. Our assessments concerning the C3 pattern are in agreement with the results reported previously by Kienle et al.;¹²

By comparing the contrasts of narrow-band (F1, F4, F5) and broadband (F6, F7, F8) images, we found that narrowing the bandwidth improved the contrast of the vascular pattern. This is due to cutting of the longer wavelengths. For example, F6 images have a lower contrast for the C1 pattern than F1 images. F6 images involve longer wavelengths than F1 images, and these longer wavelengths blur the contrast of the C1 pattern. The effects of narrowing the bandwidth on the contrast of the C2 pattern in F4 and F7 images and on the contrast of the C3 pattern in F5 and F8 images can be explained in the same way as for the C1 pattern. However, F7 and F8 band illuminations do not involve longer wavelengths, but involve shorter wavelengths than narrow-band illumination. At present, we cannot explain the mechanism responsible for this phenomenon.

Table 4 and the color images in Fig. 5 show that CIE 1976 L ^* a ^* b ^* color differences may not be suitable for evaluating the appearance of the vascular pattern in color images. On the other hand, the values of L ^* ctr agree with observations of the C1 pattern. We examined the appearance of the vascular pattern from the viewpoint of the dependence of spatial frequency on the contrast sensitivity of human vision. Lakhan and Vimal²³ and Sakata and Isono²⁴ reported that the contrast sensitivity of human vision in the luminance channel is higher than that in the color channel within a range of from 5 to 10 cycles per degree (cpd). If we assume that the observational distance is 1.4 m from the monitor, the spatial frequency of a typical C1 pattern is in the rage of 5 to 10 cpd. Referring to their results (Fig. 1 in Ref. 23 and Fig. 4 in Ref. 24), we found that the sensitivity in the luminance channel is tenfold greater than that in the color channel. The metric of L ^* ctr has no color information. Therefore it is suggested that L ^* ctr better reflects the observations of the C1 pattern than dE. In addition, C1 should be displayed as the luminance channel. To do this, we decided to use F1 and F2, and to assign both to the B and G channels. Moreover, we selected the F3 filter and assigned it to the R channel to display the C3 pattern as a bluish color. Of course other combinations of filter set and color assignment can be used by selecting the F1 filter. For example, gray-scale images can be made using only the F1 filter. However, endoscopists know from experience that information regarding the vein pattern is also important in detecting early colorectal cancer.²⁵ Therefore both the C1 and C3 patterns are important. Set A is the first trial case for clinical tests. Further studies on selection of the filter set and assignment to the color channel are required.

The tissue around the crypt orifice contains many capillaries. Therefore the pit pattern is thought to be the reverse of the capillary pattern within the tissue. F1 band images can be used to visualize the capillary pattern with high contrast and then can also be used to visualize the pit pattern. This was confirmed by a clinical test. Improvements in the appearance of the pit pattern in the BE mucosa are expected to enable detection of the SIM portion, which is associated with a high risk of developing cancer, by classifying the pit pattern. Chromoscopy also improves the contrast of the pit pattern by accumulating the dye in the crypt orifice, but if the mucosa is covered with mucus, it is difficult for the dye to accumulate. However, NBI is an optical enhancer of the contrast and is not affected by mucus. Moreover, NBI will reduce both the time and cost associated with standard chromoscopy.

The results of experiments 1 and 2 indicate that NBI with the magnified endoscope improves the contrast of the vascular pattern. On the other hand, the clinical test of small colonic polyps revealed that NBI is also useful in a wide view in which the capillary pattern cannot be recognized as individual thin vessels. The crowding of capillaries is visualized as a pattern with the appearance of brown staining. Adenomas and cancer tissue accumulate capillaries during their development. Therefore, small colonic polyps of adenomas or cancers can be seen as brown staining. Searches for small colonic polyps are usually performed under wide view. In addition, application of chromoscopy to the whole of the colon is not realistic. Therefore, NBI facilitates the finding of small colonic polyps, preventing them from being overlooked in patients with adenomas or cancers.

Hurzeler²⁶ used blue illumination to heighten the contrast of nodules, scars, and metaplasia in the bronchi. He used a blue filter with a broad bandwidth of about 250 nm. The results suggest that use of blue illumination will be useful for diagnosis in bronchoscopy. However, it is not clear whether his proposed illumination method can improve the contrast of the vascular or pit patterns; the broadband illumination used in the previous study may not be suitable for this purpose. Although the concept of the present study is similar to that of Hurzeler’s study, the method proposed here can be used to quantitatively determine the contrast effect of using narrow-band illumination and has been shown to be of clinical benefit in gastrointestinal endoscopy.

6.

Conclusions

Using five narrow-band illuminations and three broadband illuminations, we investigated the relationships between the spectral features of illumination and contrast for vascular patterns in endoscopic images. We classified the patterns into three classes (C1, C2, and C3) and quantitatively evaluated the contrast of each class. By comparing the results for narrow-band images, we found that F1 (415 nm±30 nm) is appropriate for imaging the C1 pattern; F3 (500 nm±30 nm) or F4 (540 nm±20 nm) is appropriate for imaging the C2 pattern; and F5 (600 nm±20 nm) is appropriate for imaging the C3 pattern. In addition, we found that narrowing the bandwidth can improve the contrast of the vascular pattern. Evaluating three sets of filters for color imaging, we selected a set for clinical tests. Our selected set was 415±30 nm, 445±30 nm, and 500±30 nm. Clinical tests in colonoscopy indicated that NBI aided in finding small colonic polyps. Clinical tests in Barrett’s esophageal endoscopy indicated that NBI facilitated classification of the pit pattern of the BE mucosa. These clinical benefits of NBI prevented small adenomas and cancers in the colon from being missed, enabling accurate biopsy for the portion at high risk of cancer development. We plan to conduct clinical tests in larger numbers of cases and in other diseases.