1.

Introduction

Carcinoma of the bronchi is the most common cancer and cause of cancer deaths in the world, with the highest incidence rates occurring in North America and Europe.^{1, 2, 3} Bronchial cancer has a poor prognosis, mainly due to its progressive nature and because most cancers are detected at an advanced stage.⁴ Prognosis is much more favorable if the cancer is detected and treated early, if possible at the in situ stage.^{4, 5, 6} Bronchoscopy is the only established method that allows detection, localization, and definitive histological diagnosis of endobronchial lesions. Although the resolution of bronchoscopy is superior to that of chest radiography (CRX) and computed tomography (CT), conventional white light bronchoscopy (WLB) has important diagnostic limitations for detecting early cancerous and precancerous lesions. Sato ⁷ have shown that for radiographically occult bronchial cancers in patients with positive sputum cytology, only about 30% of the lesions can be localized, with difficulty, by conventional bronchoscopy. It should be noted that the percentage of lesions missed during WLB is probably much larger, since numerous radiographically occult lesions were not detected during this study. This important clinical problem is central to the extensive research efforts performed by numerous groups to develop patient screening and cancer detection methods.^{8, 9} One promising approach to detect bronchial precancerous and early cancerous lesions during bronchoscopy is based on the imaging of the tissue autofluorescence.^{8, 10} While conventional WLB is based on the detection of minimal alterations in tissue surface structure, autofluorescence bronchoscopy (AFB) exploits the spectral and intensity contrast of the tissue autofluorescence existing between normal and pre-/early cancerous tissues.

AFB works by capitalizing on the decrease in tissue autofluorescence intensity within the green spectral region for pre- and early cancerous lesions as compared to healthy tissue under violet excitation.^{8, 11, 12} This decrease can be visualized by specific endoscopic imaging devices.¹³ In most cases, due to the three-dimensional (3-D) geometry of the bronchi, a second background image is obtained in the red part of the spectrum, to perform a contrast enhancement procedure between the contrast-bearing green and the non-contrast-bearing red images. Several endoscopic fluorescence imaging systems (Xillix LIFE system,¹⁴ Storz D-Light system,¹⁵ and Pentax 3000 system¹⁶) are based on this principle.

Numerous clinical studies have demonstrated that AFB used in addition to conventional WLB significantly increases the number of detected lesions, ^{8, 10, 14, 17, 18} increasing the sensitivity by a factor of 2. However, the higher sensitivity comes with a limited specificity.^{8, 19} Moreover, it should be noted that most of the AFB systems are likely to be suboptimal since their spectral designs are not based on the results of comprehensive studies of specific tissue autofluorescence properties. One exception is the system developed and produced by Richard Wolf GmbH. This system emerged from a systematic and comprehensive clinical study of the autofluorescence spectroscopy of normal tissues and precancerous and early cancerous lesions of the mucous membrane lining the tracheo-bronchial tree.^{11, 20} The group in Lausanne demonstrated that the highest contrasts between healthy and (pre-) malignant tissues are observed with excitation wavelengths around $405\mathrm{nm}$ . In addition, they observed that, at this excitation wavelength, a strong decrease of fluorescence intensity exists for these lesions in the green (below $590\mathrm{nm}$ ) region of the spectrum, whereas this decrease is much smaller but still present in the red part of the spectrum (beyond $590\mathrm{nm}$ ). Typical spectra from healthy and dysplastic bronchial mucosa excited at $405\mathrm{nm}$ are shown in Fig. 1 . Since a small decrease is also observed in the red part of the AF spectrum, this group suggested a possible method to improve the performances of fluorescence bronchoscopy: using additional red backscattered light as a background instead of the red autofluorescence. According to Zellweger, ¹¹ using an optimized amount of red backscattered light should increase the contrast between healthy and pre-/early malignant tissue by a factor of 2. Moreover, the background image created by the backscattered red light enhances the brightness and consequently the quality of the endoscopic images. Last, the choice of backscattered red light appears to be a sensible spectral choice, as it is minimally influenced by changes of tissue properties, textures, and structures. It should be noted that a similar approach using backscattered red and near-infrared light was reported by Zeng ²¹ for autofluorescence imaging in the gastrointestinal tract.

Fig. 1

Autofluorescence spectra from normal (circles) bronchial tissue and dysplasia/CIS (triangles) (adapted from Ref. 11). These spectra were obtained using an excitation of $405\mathrm{nm}$ , $\mathrm{FWHM}=10\mathrm{nm}$ with an optical fiber–based spectrofluorometer. The healthy tissue spectrum is dominated by a peak around $500\mathrm{nm}$ , i.e., in the green part of the spectrum, while the spectrum measured on dysplastic tissue shows equal intensities in the green and the red part of the visible spectrum. The dotted line separates the green, i.e., short-wavelength, region and the red, i.e., long-wavelength, region used to generate the foreground and background autofluorescence images, respectively.

The impact of these spectral improvements, according to our knowledge, has not previously been assessed quantitatively with imaging systems. Therefore, we have designed an endoscopic fluorescence imaging system that includes the features of violet excitation around $405\mathrm{nm}$ and offers the option to add red light to the excitation. This system was used for the clinical study reported here to confirm the results from Zellweger ¹¹ and to assess the potential of additional backscattered red light for improved contrast between healthy and (pre-)malignant tissue.

2.

Materials and Methods

2.1.

Instrumentation

We used a modified endoscopic fluorescence imaging (EFI) system developed in collaboration with the company Richard Wolf Endoskope GmbH, Knittlingen, Germany. The system consisted of a modified endoscopic light source, a filtered 3 CCD RGB camera (5137 Combilight PDD and EFI 5507 camera), and a dedicated bronchofibroscope. This experimental system offered the unique advantage of three different illumination/excitation modes that could be selected during the bronchoscopy using a simple foot switch. The light source contained an infrared (IR)-filtered 300-W Xe lamp and was equipped with a flip-flop filter holder allowing for change between white light illumination for conventional endoscopy and two different violet light excitations for autofluorescence endoscopy. The two violet fluorescence excitation filters used in our study were: a violet bandpass filter (pure violet: PV) with a transmission $⩾95\%$ between 395 and $460\mathrm{nm}$ , and a double bandpass filter (violet plus red light: $\mathrm{V}+\mathrm{R}$ ) with a transmission $⩾$ 90% between 390 to $470\mathrm{nm}$ and a second weak (8%) transmission between $650\mathrm{nm}$ to $680\mathrm{nm}$ . The transmission spectra of these filters are shown in Fig. 2 . The slight spectral differences in the blue-violet bands were conditional of manufacturing but do not affect the excitation properties. Both filters transmit the same light energy, although the transmission band of the $\mathrm{V}+\mathrm{R}$ filter is slightly broader than that of the PV filter. This is due to the lower transmission of the $\mathrm{V}+\mathrm{R}$ filter relative to the PV filter. Both filters were used alternately during the endoscopic exams. The illumination/excitation light was transmitted from the light source to the endoscope via a liquid light guide (Type 4070.253, Richard Wolf GmbH, Knittlingen, Germany).

Fig. 2

Transmission spectra of the excitation filters. The transmission spectrum of the pure violet (PV) bandpass filter is shown on the left. The graph on the right shows the transmission characteristics of the double bandpass filter used for violet excitation in combination with the detection of the backscattered red light $(\mathrm{V}+\mathrm{R})$ .

As can be seen in Fig. 3 , the camera objective contained a 490-nm long-pass filter that was optimized to reject all violet excitation light. A communication cable connected the camera controller and the light source. When the light source was used in the conventional white light mode, the camera employed the standard video red, green, and blue color (RGB) mode. If the light source was used in one of the fluorescence excitation modes, the signal of the camera’s blue channel was suppressed. Consequently, images taken in the fluorescence mode contained only the red and green image color information necessary to visualize the spectral contrast between healthy and precancerous tissue, as described in Sec. 1. The gamma factor of the camera-video unit was 0.71.

Fig. 3

Block diagram of the AF bronchoscopy system.

The camera head was clipped to a flexible bronchofibroscope. Unlike conventional bronchofibroscopes, this device was equipped with illumination fiber bundles offering a high UV transmittance. The available total illumination power at the distal end of the bronchoscope was typically $120\mathrm{mW}$ with both excitation filters. This power was high enough to enable a real-time detection of the autofluorescence/reflectance images in the entire tracheo-bronchial tree at the video frequency. In addition, this camera enabled the user to perform a “dual” automatic white/color balance, as explained below. A block diagram of the system is shown in Fig. 3.

3.

Results

3.3.

Illustrations of WLB and AFB Imaging

The following images show true positive lesions examined with WLB, PV fluorescence excitation, and $\mathrm{V}+\mathrm{R}$ fluorescence excitation.

Figure 4 shows a mild dysplasia presented by a 62-year-old male patient observed with conventional WLB and AFB with PV and $\mathrm{V}+\mathrm{R}$ fluorescence excitation, respectively. This patient was known to suffer from a squamous cell carcinoma in the upper right bronchus and underwent pretherapeutic bronchoscopy for treatment by photodynamic therapy. The images show the division between the middle and lower lobe. The spur was slightly suspicious with WLB due to thickening and reddening.

Fig. 4

Mild dysplasia. The lesion on the spur is hardly suspicious with WLB but appears as a sprawled brownish area with AFB using pure violet excitation light. When violet excitation plus additional backscattered red light is used, the lesion appears as a marked reddish area on the greenish surrounding tissue. Moreover 3-D perception is enhanced in the latter image.

AFB exhibited a sprawled brownish area on the olive-green healthy tissue background under pure violet excitation. The same area is clearly visible as a bright red zone on a clear green background when backscattered red light is used in addition to the violet excitation.

These images illustrate the improved detectability and demarcation of the extension of superficial bronchial lesions if AFB is used instead of WLB. Moreover, the use of the $\mathrm{V}+\mathrm{R}$ excitation filter clearly allows us to generate much better images than the PV filter. In addition, the images obtained with the $\mathrm{V}+\mathrm{R}$ excitation are much brighter and less noisy than their counterparts obtained with PV excitation only. This is due to an enhancement of the chromatic contrast between the lesion and the surrounding healthy tissue.

Figure 5 illustrates the effect of backscattered red light on the background image in the red color channel. The AF images from a case of severe dysplasia observed with PV and $\mathrm{V}+\mathrm{R}$ excitation were separated into their green (left image) and red (right image) channels, respectively, and are presented in gray scale. The lesion is visible as a dark, sharp circumscribed area on the base of the spur. In the case of PV excitation [Fig. 5a], both the green and the red channel images show an intensity decrease on the site of the lesion. In the case of $\mathrm{V}+\mathrm{R}$ excitation [Fig. 5b], the red channel image exhibits a nearly homogenous intensity distribution, while the green channel image shows an intensity decrease on the site of the lesion. It should be noted that the spur was not classified as suspicious during WLB.

Fig. 5

Influence of the backscattered red light on the image background. These images present the same lesion as in Fig. 4, but with the fluorescence images split into their green and red color channels. They are presented in grayscale mode. Addition of backscattered red light to the violet fluorescence excitation light (a) results in a nearly homogenous background in the red channel image. The corresponding image obtained with PV excitation (b) shows demarcated zones of decreased intensity, influencing the contrast between diseased and healthy tissue areas.

3.4.

Image Analysis

The mean normalized red-to-green ratios computed from image analysis of fluorescence positive areas and the fluorescence negative healthy reference zones are shown in Fig. 6 for both fluorescence excitation modes. Their values and standard deviations are shown for the different histopathologic classes—normal tissue (squares), reactive changes (triangles), and (pre-)neoplastic changes (diamonds). The circles depict the mean ratios from the fluorescence negative healthy reference zones.

Fig. 6

Mean values of normalized ratios for both pure violet and $\mathrm{violet}+\mathrm{red}$ excitation. This graphic depicts the mean values of the normalized ratios for pure violet excitation (data on the left) and violet excitation with backscattered red light (data on the right). Circles denote values for healthy and fluorescence-negative tissue, squares denote healthy but fluorescence-positive tissue, triangles denote reactive changes (inflammation and metaplasia), and diamonds denote (pre-)neoplastic changes (dysplasia and carcinoma). The number of cases, $n$ , is given for each histopathologic class.

Table 3

Number of true positive (TP) and false positive (FP) as well as the true negative (TN) and false negative (FN) cases for WLB and AFB with backscattered red light.

	TP	FN	FP	TN
AFB	16	0	19	2
WL	12	4	16	5

In the case of pure violet excitation, the mean ratios were comparable for false positive normal $({\mathrm{mean}}_{\mathrm{normal}}=1.66\pm 0.39)$ , reactive $({\mathrm{mean}}_{\mathrm{reactive}}=1.75\pm 0.12)$ , and (pre-)neoplastic changes $({\mathrm{mean}}_{\left(\mathrm{pre}\text{-}\right)\mathrm{neoplastic}}=1.74\pm 0.43)$ . It should be noted that the mean value of the normalized ratio for the fluorescence-negative healthy reference zone was ${\mathrm{mean}}_{\mathrm{reference}}=1.03\pm 0.05$ . Adding backscattered red light to the violet excitation increased the corresponding mean values by a factor of about 2 for the normal and reactive changes classes ( ${\mathrm{mean}}_{\mathrm{normal}}=3.65\pm 0.43$ , ${\mathrm{mean}}_{\mathrm{reactive}}=4.04\pm 0.48$ ). For the (pre-) malignant lesions, this factor exceeded the value of 2.7 ( ${\mathrm{mean}}_{\left(\mathrm{pre}\text{-}\right)\mathrm{neoplastic}}$ for violet plus $\mathrm{red}=4.93\pm 2.25$ ). The mean value of the normalized ratio for the fluorescence-negative healthy reference zone remained in the range of 1 $({\mathrm{mean}}_{\mathrm{reference}}=1.10\pm 0.16)$ .

The signal-to-noise ratio (SNR), defined as the average pixel intensity of the analyzed tissue sites divided by its standard deviation, was by a factor of about 2 higher for the $\mathrm{V}+\mathrm{R}$ excitation $(\mathrm{SNR}=8.8)$ than for the pure violet excitation $(\mathrm{SNR}=4.5)$ .

The ROC comparative analysis was based on a TPF of 0.66 (66%). The corresponding FPFs for the PV method were 0.61 (fluorescence positive FP, but healthy) and 0.95 (reactive tissue changes). The corresponding values for the $\mathrm{V}+\mathrm{R}$ method were 0.26 (healthy) and 0.6 (reactive changes).

4.

Discussion

AFB has proven to be an efficient and useful tool in the detection of premalignant lesions and early cancers in the bronchi. In this study, we present an optimized AFB imaging system offering superior color contrast between healthy tissue and lesions.

One of the major problems encountered by most commercially available AFB systems is the poor image quality. This is due to the limited tissue autofluorescence intensity, combined with the limited light collection efficiency, throughput, and physical sensitivity of the imaging system.²⁴ Our system employed a filtered Xe-lamp light source in combination with a fibroscope specially designed for autofluorescence detection with an excitation wavelength around $405\mathrm{nm}$ . Used in combination with a 3 CCD endoscopic color camera, we obtained noise-reduced fluorescence images at video frequency, i.e., about 25 frames per second for PAL video, bright enough for a proper navigation in the tracheo-bronchial tree. This enabled the physician to perform the entire bronchoscopy in the autofluorescence mode instead of examining the bronchial tree in the conventional white light mode, switching to the autofluorescence mode only at specific sites.

Due to the 3-D geometry of the organ, autofluoresence detection in the bronchi requires the detection of both a background and a contrast-bearing foreground image. Such a background image can be created by backscattered light detected in one or several different color channels.

Profio²⁵ and Leonhard¹⁵ reported the detection of a small amount of backscattered blue excitation light to optimize the quality of autofluorescence detection in the bronchi. Indeed, the use of blue/violet backscattered light helps to reveal the texture of the bronchi, but it is not ideal for correcting the autofluorescence images detected in the green spectral domain. Consequently, following the spectrofluorometric study of the bronchial autofluorescence performed by our group,¹¹ we investigated the use of backscattered red light for the optimization of our imaging system.

We demonstrated in the study presented in this paper that the combination of violet light induced tissue autofluorescence with backscattered red light increases the red-to-green color ratios between normal and (pre-)neoplastic tissue areas in the AFB images by a factor of about 2.7 compared to violet-light-induced fluorescence alone. This factor corresponds to the autofluorescence intensity ratio measured spectroscopically in the red part of the spectrum, i.e., between 600 and $800\mathrm{nm}$ , between healthy and malignant tissues reported by Zellweger ¹¹ This twofold to threefold increase of the red-to-green ratio is responsible for the improved chromatic contrast perceived in the $\mathrm{V}+\mathrm{R}$ fluorescence images. Indeed, the bright red appearance of fluorescence positive lesions excited with violet light and backscattered red light are more easily perceived on the green background characteristic of healthy tissue than the brownish-olive colored lesions observed under pure violet excitation. Moreover, the red background image created by the backscattered red light dramatically improves the SNR, allowing better navigation and orientation for the endoscopist in the autofluorescence mode. In addition, the use of the $\mathrm{V}+\mathrm{R}$ excitation filter strongly suggests an increase in specificity, as indicated by Fig. 6. The overlap between the distribution of the normalized ratio measured in true positive and false positive sites is much smaller with the V+P filter than with the PV filter, indicating a more pronounced demarcation. This is confirmed by the results from the ROC analysis, which showed that for a TPF of 0.66, the corresponding FPF was lower by a factor of about 2 for the $\mathrm{V}+\mathrm{R}$ excitation than for the PV excitation.

Normal, reactive changes and (pre-)neoplastic fluorescence positive sites showed no significant differences in their normalized red-to-green ratios under PV excitation (CI 95%, $p>0.283$ ). When additional backscattered red light was used, a significant difference could be observed between fluorescence positive but histopathologically normal sites (false positives, FP) and fluorescence positive (pre-)neoplasia (true positives, TP) (CI 95%, $p=0.023$ ). A much weaker difference (CI 80%, $p=0.098$ ) was observed between the ratios computed from the FP reactive changes and the TP (pre-)neoplasia excited with the $\mathrm{V}+\mathrm{R}$ excitation.

The ratios computed from the healthy fluorescence negative controls were all significantly different (CI 95%) than the ratios computed from the fluorescence true positive (pre-)neoplastic and the false positive healthy or reactive sites for both the PV and $\mathrm{V}+\mathrm{R}$ filter. However, the p-values computed for the fluorescence negative controls and the fluorescence positive normal, reactive changes and (pre-)neoplastic lesions were by one to seven orders of magnitude lower for the $\mathrm{V}+\mathrm{R}$ excitation [normal $p=4.21\times {10}^{-4}$ , reactive changes $p=1.24\times {10}^{-5}$ , and (pre-)neoplasia $p=2.99\times {10}^{-6}$ ] than for the PV excitation (normal $p=1.31\times {10}^{-3}$ , reactive changes $p=1.45\times {10}^{-5}$ , and (pre-)neoplasia $p=2.79\times {10}^{-4}$ ). Moreover, it is worth noting that the SNR improved by a factor of about 2 when $\mathrm{V}+\mathrm{R}$ excitation was used. Therefore, the results presented in this paper indicate that the addition of backscattered red light provides an opportunity to better discriminate lesions and healthy tissue. We would like to emphasize that the results obtained in this study do not allow us to compute the sensitivity or specificity of AFB, since no random biopsies were taken.

The issue of false positive tissue sites has already been reported by other authors using commercial systems based on violet autofluorescence excitation. ^{26, 27, 28, 29} Several reasons may account for the high number of false positives. It is likely that some type of somehow atypical but noncancerous tissues appear positive under AFB, including scar tissue,²⁸ inflammation, and bruised mucosa. Therefore, the patient inclusion criteria of our study (patients with known endobronchial lesions, lung cancer follow-up patients, or high-risk patients suspect for endobronchial lesions) can influence the false positive rates. In the study presented here, about 45% of all patients exhibiting fluorescence positive sites correspond to the latter two groups. Clinical studies including these types of patient groups frequently report increased false positive rates.^{30, 31, 32} Another source of false positives in clinical studies is related to the procedure of taking a biopsy. Whenever possible, biopsies were taken under fluorescence light excitation to assure that the tissue sample was taken from the fluorescence positive area. Nevertheless, it cannot be excluded that a certain number of tissue samples were not taken from the fluorescence positive regions, but from neighboring nonfluorescent tissue. Last, histopathologic examination is deemed as the “gold standard” for comparison with bronchoscopic findings in nearly all clinical studies reporting on autofluorescence bronchoscopy. Nevertheless, some authors report on significant interobserver variability in histopathologic reporting.^{33, 34} In a retrospective study of 343 bronchial tissue samples from AFB, Venmans ³⁴ revealed interobserver variability by a factor of 2 in the number of diagnosed preinvasive bronchial neoplasias. Thus, the interobserver variability in the histologic analysis of preinvasive bronchial tissue changes has an influence on the results of conventional bronchoscopy as well as AFB. The tissue samples obtained in our study were analyzed on a single-view basis only, i.e., each biopsy was analyzed by a single pathologist. It is therefore very likely that a certain number of AFB positive but histologically reported negative, i.e., AFB false positive cases, are in reality AFB true positive results. This would dramatically change the values of sensitivity and PPV. Independent revision of the samples by two or more histopathologists might reduce such errors. The distribution of the normalized ratio for the healthy tissue is of particular interest in this context. Indeed, this distribution seems to be bimodal, the true negative sites being closely grouped around the value of one, whereas the normalized ratios of the false positive site are grouped at much larger values. This effect is particularly visible with the $\mathrm{V}+\mathrm{R}$ excitation. It should be noted that the two sources of possible incorrect histopathologic assignations, i.e., the precision of tissue sampling and the interobserver variability, could explain the bimodal distribution of the ratios corresponding to the healthy tissues.

Large interpatient fluctuations of the normalized red-to-green ratios within the histologic classes were observed with both excitation modes. These fluctuations were more marked when backscattered red light was added to the violet excitation. This is indeed a drawback of the method. Analysis of the origins of these variations will lead to a better understanding of this phenomenon. Interpatient fluctuations in the AF were reported from several spectrofluorometric studies in the bronchi²⁰ and in other organs.^{21, 35} Since an important contribution to the fluctuations of the tissue autofluorescence intensity can be attributed to the inter- and intrapatient variations in fluorochrome concentrations and epithelium thickness, these fluctuations are more or less proportional in the green and red parts of the spectrum (intrinsic fluctuations).^{11, 20, 36} However, this proportionality is no longer present if red backscattered light is detected instead of red autofluorescence. It should be noted that a similar phenomenon is present regarding the intensity of the excitation light available for the excitation of the fluorochromes. Indeed, our fluorescence imaging system uses an excitation band around $410\mathrm{nm}$ , corresponding to one absorption peak of hemoglobin. The main alterations in the local intensity of violet light available for excitation of the tissue autofluorescence are therefore related to the presence and extent of blood and blood-containing structures, i.e., for instance, the density, size, and depth of the blood vessels underlying the bronchial epithelium. These factors can vary markedly from one site to another. In the case of pure violet excitation, both the green and the red autofluorescence images are detected. A decrease in the violet excitation light due to the abovementioned local blood-containing structures therefore results in a nearly proportional decrease in the green as well as in the red part of the spectrum. Consequently, in the case of pure violet excitation, the inter- and intrapatient variations of the red-to-green ratios are minimally influenced by the blood-related fluctuations of the excitation light intensity. When backscattered red light is used in addition to the violet excitation, only the autofluorescence in the green image will be influenced by the absorption of the excitation light.

The enhanced chromatic contrast between a lesion and its healthy surrounding tissue allows a precise demarcation of the lesion’s margins. This makes AFB an important tool, not only for diagnostic but also for pretherapeutic and presurgical bronchoscopy, where the determination of tumor margins is crucial for the planning and outcome of the treatment or surgery.

In summary, the results presented in this paper demonstrate that the additional detection of backscattered red light in violet-light-induced AFB has the potential to significantly increase the chromatic contrast between healthy and abnormal bronchial tissues. Moreover, the additional light markedly reduces the SNR, thus improving the image quality. This is of high interest for the development of future AFB systems. In the study presented here, each bronchoscopy was performed with fiber optic endoscopes equipped with an endoscopic camera clipped to the endoscope’s eyepiece. However, the use of flexible videobronchoscopes has increased dramatically in recent years. It should be noted that autofluoresence videobronchoscopy systems have been commercially available for several years. The implementation of the principle of $\mathrm{V}+\mathrm{R}$ detection has the potential to further improve the image quality and healthy-to-lesion contrast in those systems.