2.1.

Patient Selection

Twenty-three patients (18 men and 5 women) with a mean age of $65.0\text{years}$ (range $46\text{to}81\text{years}$ ) who are suspected of primary or recurrent bladder cancer were enrolled for this clinical study at the Singapore General Hospital with no exclusion criteria, following their informed consent. Fifty-eight hypericin fluorescence cystoscopic images were obtained from these patients, together with their corresponding histopathological status from biopsy specimens taken from the same imaging site. The diagnosis based on histopathology was used as the gold standard for classifying the images into the various grades of cancer. This study was reviewed and approved by the Ethics Committee of the Singapore General Hospital and National Cancer Centre Singapore.

2.2.

Fluorescence Cystoscopy

Patients were administered with hypericin via intravesical instillation into the bladder, followed by an in situ incubation period of $2\mathrm{h}$ in the bladder. The bladder instillation solution was prepared from $16\mathrm{mM}$ hypericin dissolved in absolute ethanol, diluted 1,000-fold in a 1% plasma protein solution in buffered saline and filtered with a sterile membrane. Aliquots of $25\mathrm{ml}$ were then kept frozen until use and $25\mathrm{ml}$ saline was added to obtain a $50\text{-}\mathrm{ml}$ instillation solution containing $8\mu \mathrm{M}$ hypericin. A preliminary study on patients with papillary superficial Ta disease had shown that this was the preferred concentration and volume for adequate hypericin fluorescence.¹⁶ The patients were advised to change their position at regular intervals so that the hypericin was well distributed in the bladder tissue. After the incubation period, the bladder was emptied, and fluorescence cystoscopy was performed swiftly on the patient’s bladder under excitation with blue light $\left(370\text{to}450\mathrm{nm}\right)$ to minimize any possibilities of hypericin photobleaching. Biopsies of suspicious red fluorescent regions and surrounding nonfluorescent regions were taken from the imaged tissue sites and sent for routine histopathology to obtain the tissue histological classification for the corresponding fluorescence images.

2.3.

Instrumentation

The digitized fluorescence endoscopic imaging system used in this study was previously described in detail, and its schematic is shown in Fig. 1 .²⁶ Briefly, the system consists of an illumination console, a fluorescence detection unit, a video displaying and recording unit, and a computer system for image acquisition, display and processing. A $100\text{-}\mathrm{W}$ xenon short arc lamp (D-Light AF system, Karl Storz, Germany) was used for both the white-light illumination and the hypericin fluorescence excitation when filtered by a bandpass filter $\left(370\text{to}450\mathrm{nm}\right)$ . The irradiance of the blue light at the cystoscope tip was approximately $50\mathrm{mW}∕{\mathrm{cm}}^2$ . Both illumination and observation of tissue of interest were achieved via a modified cystoscope integrated with an optical long-pass filter in the cystoscopic eyepiece with cutoff wavelength at $470\mathrm{nm}$ that transmits only 8% of the diffusely backscattered excitation blue light from the tissue but over 98% of light in the $470\text{to}800\text{-}\mathrm{nm}$ range. This minimizes the backscattered blue excitation light while allowing the red hypericin fluorescence to pass through the cystoscope efficiently.²⁷ The fluorescence images were detected by a sensitive three-chip color CCD video camera with a pixel resolution of $768\times 592$ (Tricam SL-PDD, Karl Storz, Tuttlingen, Germany), connected to the modified cystoscope with its RGB video outputs fed simultaneously to a frame grabber (Matrox Genesis-LC) at a rate of 25 frames per second for capturing and digitizing the images. The digitized images were stored for offline image processing, and the processed image can be displayed on the computer screen in near real time for further analysis and diagnosis.

Fig. 1

Schematic diagram of the digitized fluorescence endoscopy imaging system for early cancer diagnosis. WL mode—white light illumination mode; HYP mode—hypericin fluorescence excitation mode.

2.4.

Image Analysis

An image processing routine was developed from the Olympus MicroImage 4.0 (Media Cybernetics) software to carry out the various processing, including contrast enhancement, histogram thresholding, and segmentation, prior to the colorimetric analysis, as shown in Fig. 2 . The image contrast was first enhanced by histogram equalization, i.e., stretching the image histogram to fill all 256 levels of the $8\text{-}\text{bit}$ image, after which the histogram for the hue was extracted for the entire image. This global histogram will typically be bimodal, with two peaks in the hue, one attributed to the normal blue background and the other peak attributed to the red fluorescence from the suspicious lesion. A threshold between the two peaks was then determined by finding the minimum point between the two modes to define the range of hue values to be considered at suspicious against the rest of the normal background. Hue values above the threshold (red region) were considered in the area of interest (AOI). An AOI based on such hue segmentation of lesion area in the cystoscopic images was then defined to segment the suspicious lesion area out from the rest of the background image. The fluorescence intensities of the red, blue, and green channels of the fluorescence images within the AOI were then extracted and quantified. We note that no saturation in the intensity was observed for any of the three color channels, which may affect the results. Intensity ratios were calculated from the three colorimetric parameters to get the red-to-green intensity ratio $(I_R∕I_G)$ and the red-to-blue intensity ratio $(I_R∕I_B)$ within the AOI. Statistical averages of the intensity ratios for all the patient images under each histological classification were then calculated and compared with histopathological status of the imaged sites to ascertain any correlation. The possible changes in the intensity ratios with the observation distance between the tip of the endoscope and the tissue had previously been reported by our group.^{22, 23} Despite the decreased fluorescence intensity with increasing imaging distance, the intensity ratios $I_R∕I_G$ and $I_R∕I_B$ remained nearly consistent in the normal tissue or in the tumor. We noted that the intensity ratios were also independent of the observation geometry and the fluctuation of the excitation irradiance, since all three color channels depend on the distance, angles of the irradiation and observation, and the tissue optical properties in a similar manner. The intensity ratios thus form a robust diagnostic algorithm for cancer grading in this study.

Fig. 2

Schematics of the algorithm used in fluorescence image processing to extract the three colorimetric parameters and hence obtain the two diagnostic intensity ratios.

2.5.

Statistical and Histopathological Analysis

In assessing the pathological status, the histopathological results were regarded as the gold standard. The tumors were graded by a central pathologist unaware of origin, using the international systems for urothelial neoplasia. In this study, hypervascularized, erythematous, erosive, or edematous areas of the mucosa were considered nonspecific inflammation of the bladder mucosa. Normal, inflammation, and hyperplasias were classified as benign conditions. All three grades of invasive TCC, i.e., grades I, II, and III were classified as malignant pathological changes. The unpaired two-sided Student’s $t$ -test was performed on the fluorescence intensity ratios $I_R∕I_G$ and $I_R∕I_B$ to determine whether they were statistically significant for discriminating malignant from benign tissue and the one-way ANOVA test with one-way analysis of variance was performed on the fluorescence intensity ratios $I_R∕I_G$ and $I_R∕I_B$ to determine whether they were statistically significant for discriminating between the multiple groups, i.e., normal, inflammatory hyperplasia, and three different grades of TCC. The sensitivity and specificity of using hypericin fluorescence for grading the bladder cancer in situ were also determined as defined here⁴:

3.1.

Patient Classification and Side Effects

Overall, 58 cystoscopic images from 23 patients were taken from various sites of the bladder in which, after histopathological confirmation, 10 were found to be normal, 12 hyperplasias with inflammation, 11 grade I TCC, 13 grade II TCC, and 12 grade III TCC. The breakdown of histopathological status of the cystoscopic images is summarized in Table 1 . All 23 patients involved in this study were subjected to intravesical instillation of hypericin. One major advantage of intravesical instillation of hypericin is that it is topical and therefore avoids the intravenous route and systemic side effects such as cutaneous adverse photoreactions due to photosensitivity. There were no reports of skin photosensitization or other side effects found in patients who have undergone hypericin fluorescence cystoscopy. There are also no reports of phototoxicity and photoreaction using topical hypericin in the bladder as well as other clinically significant short-term side effects such as urinary tract infection.

Table 1

Histopathological classification of the 58 cystoscopic images used in this study.

3.2.

Fluorescence Cystoscopic Imaging and Tumor Detection

In this study, we assessed the potential of applying image analysis on hypericin fluorescence to grade the bladder cancer in situ based on the 58 fluorescent cystoscopic images from the 23 patients. Although the qualitative description of microscopic hypericin fluorescence in ex vivo bladder tissue of TCC as well as macroscopic hypericin fluorescence in the detection of CIS had been previously reported in the literature,^{16, 17, 18} the quantification of hypericin fluorescence for grading of bladder cancer has not yet been reported, to our knowledge. We present our preliminary study with the aim to correlate this fluorescence with the different pathological grades of bladder cancer.

Our previous and current clinical studies have demonstrated that the digitized fluorescence endoscopic system used in this study has the capability of acquiring high-quality images in several milliseconds so as to minimize the effect of hypericin photobleaching and shorten the required time for holding the endoscope steadily. Figure 3 shows representative fluorescence cystoscopic images acquired from normal, inflammatory, and each of the three grades of TCC lesions before any image processing. The fluorescence cystoscopic images showed that suspicious lesions display bright reddish fluorescence, while the surrounding normal bladder mucosa exhibit blue background with no red fluorescence. It was observed that apart from malignant lesions, benign lesions with chronic inflammation also showed reddish fluorescence, i.e., false-positive fluorescence. This may be due to a host of different factors, such as physiological changes in the inflammatory site that also favor the transport and accumulation of hypericin there, the presence of porphyrin and its derivatives from bacteria associated with inflammation at the sites, etc. However, we noticed that the benign lesions with chronic inflammation generally showed milder red fluorescence compared to the TCC cases and that, visually, the fluorescence-positive malignant lesions had significantly higher red-to-blue fluorescence contrast to the surrounding normal tissue as compared to the fluorescence-positive benign lesions.

Fig. 3

Representative cystoscopic images of hypericin fluorescence in (a) normal region of the bladder, (b) inflammatory region of the bladder, (c) Grade I TCC, (d) grade II TCC, and (e) grade III TCC showing the differences in fluorescence intensities.

The images were processed to extract their various colorimetric intensities. After image processing, the intensity of each of the three constituent colors of the fluorescence images, i.e., the red hypericin fluorescence intensity $I_R$ , green autofluorescence $I_G$ , and the backscattered blue excitation light intensity $I_B$ , of the tissue sites of interest was extracted and quantified. The mean intensity values of the $8\text{-}\text{bit}$ pixel data, i.e., 0 to 255 for each of the three constituent color channels for benign and malignant tissue are shown in Table 2 for comparison. These quantitative intensities concur with our qualitative visual observations that the red fluorescence intensity $I_R$ of malignant tissue (all three grades of TCC) is stronger than that of benign (normal, inflammation, and hyperplasia) tissue. In addition, it is also observed that the diffusely backscattered blue excitation light intensity $I_B$ and the green autofluorescence $I_G$ of malignant tissue is less than that of benign tissue.

Table 2

Comparison of different colorimetric parameters extracted from cystoscopic images of benign and malignant bladder tissue.

Combining all three (red, green, and blue) elements of the clinically valuable fluorescence information, we have chosen both the red-to-green $(I_R∕I_G)$ and red-to-blue $(I_R∕I_B)$ intensity ratios as appropriate diagnostic intensity ratios to further enhance the colorimetric differences between benign and the different grades of malignant tissue for fluorescence diagnosis. These intensity ratios may provide an indication to the changes in tissue architecture and fluorophore contributions in malignancies. From the intensity ratios, it was found that both the mean $I_R∕I_G$ and $I_R∕I_B$ values for malignant tissues are generally greater than those of benign tissues. The mean ratio value of $I_R∕I_G$ in malignant tissue is $4.48\pm 1.75$ , which is significantly higher compared to the mean value of $2.63\pm 1.92$ in benign tissue (unpaired Student’s $t$ -test, $P=0.00018<0.0005$ ). Similarly, the mean $I_R∕I_B$ value of $1.07\pm 0.62$ in malignant tissue is also significantly higher than that of $0.56\pm 0.37$ in benign tissue (unpaired Student’s $t$ -test, $P=0.00107<0.002$ ).

An interpretation of the observed intensity changes for each of the three color channels with malignancy requires an understanding of the changes in tissue architecture and fluorophore contribution that comes along with carcinoma progression. With the $380\text{to}450\text{-}\mathrm{nm}$ excitation light, the enhanced red fluorescence observed in malignant tissue is mainly due to the photosensitizer-induced fluorescence associated with the selective accumulation of hypericin in malignant lesions, although the endogenous autofluorescence or the presence of porphyrin and its derivatives may have also contributed to the observed red fluorescence, as in cases of bacterial associated inflammation. The green intensity typically represents the intrinsic tissue autofluorescence, and blue intensity largely represents the amount of backscattered light and, to a smaller extent, the blue tissue autofluorescence.

The increase in the red fluorescence intensity for all three grades of malignant TCC over that of benign tissue can be attributed to the greater selective accumulation of hypericin in malignant tissue, possibly due to enhance permeation of the photosensitizer from the tumor vasculature into the tumor interstitial as carried by lipoproteins, coupled with a compromised lymphatic drainage that retards the efflux of hypericin out of tumors. The decrease in green and blue tissue autofluorescence in malignant tissues may be due to the thickening of the epithelial layer, increasing blood volume in the tumor from increase in microvascular density, and decrease in fluorescence quantum yields of endogenous fluorophores, e.g., collagen, elastin, NADH, and flavin. It is known that the fluorescent co-enzymes NADH and FAD, which are important in aerobic metabolism, are normally found in large amounts in normal cells but are absent or expressed in low amounts in anaerobic cancer cells. Besides tissue autofluorescence, the decrease in the diffusely backscattered blue excitation light may be attributed to the increased hemoglobin absorption by a larger blood volume in malignant lesions.²⁸

3.3.

Correlation of Colorimetric Intensities with Cancer Grades

We took the analysis of the intensity ratios a step further and observed that differences in both the $I_R∕I_G$ and $I_R∕I_B$ intensity ratios were also observed within different grades of malignancies. In fact, these intensity ratios generally exhibit a correlation with cancer malignancy. The statistical means of both intensity ratios decrease with higher grade of bladder cancer. Figure 4 shows the variation of the mean intensity ratios $I_R∕I_G$ and $I_R∕I_B$ with the three histopathological grades of bladder cancer available in this study in the form of a histogram. The mean intensity ratios for normal and inflammatory conditions are also shown for comparison. From the histogram, it can be seen that a correlation exists between the statistical means of both $I_R∕I_G$ and $I_R∕I_B$ and the histopathological grade of the bladder cancer tissue and that, specifically, a decrease in both $I_R∕I_G$ and $I_R∕I_B$ was observed with increasing level of carcinoma differentiation. Using the one-way ANOVA test with one-way analysis of variance for more than three groups of samples, we have determined that statistical differences between five groups of samples, i.e., normal, inflammation, and three different grades of malignant bladder cancer are significant for both the red-to-blue intensity ratio $(P<0.05)$ and the red-to-green intensity ratio $(P<0.05)$ . In fact, the $P$ -value for the red-to-blue intensity ratio is 0.0003, and the $P$ -value for the red-to-green intensity ratio is $<0.0001$ . Apart from the observed correlation between the intensity ratios and the histopathological grade of cancer tissue, it is also observed that both the mean intensity ratios of the highest grade of TCC overlap with the hyperplastic inflammatory lesions. As discussed earlier, the fluorescence observed in these inflammatory lesions could be due to many different factors, such as similar physiological changes that are taking place in the inflammatory sites as compared to the carcinoma sites that favor the selective accumulation of hypericin. Such similarities may account for the overlap in the mean intensity ratios, thus resulting in false positives, and may present further implications to the performance of this technique in grading TCC, as discussed in more detail shortly. Nonetheless, our results still highlight the potential of the two diagnostic intensity ratios based on $I_R∕I_G$ and $I_R∕I_B$ to discriminate between the different grades of malignant tissue.

Fig. 4

Comparison of $I_R∕I_G$ and $I_R∕I_B$ between the three different grades of bladder cancer and benign tissues (normal and inflammation). The number above each bar indicates the sample size for each category of pathology in this study.

Following our earlier statement that an increase in the intensity ratios (due to the higher red intensity and lower autofluorescence intensity) indicates a higher level of hypericin in the tumor, these results seem to intuitively suggest that higher-grade bladder tumors either present a smaller uptake of hypericin during intravesical instillation or a more-effective efflux of hypericin out of the tumors after instillation. However, we attempt to briefly account for the observed correlation based on the alterations in the extracellular matrix in tumors affecting the paracellular transport processes of hypericin through the bladder mucosa and the tissue optics of the excitation light. Previous studies have suggested that the alterations in expression of transmembrane glycoproteins such as cadherins and integrins that mediate the cell-cell and the cell-extracellular matrix adhesion result in modifications in the intercellular adhesion properties of cancer cells during carcinogenesis.²⁹ Loss of intercellular adhesion during carcinogenesis allows the malignant cells to degrade the cellular matrix for invasion and metastasis according to their aggressiveness. This suggests that the modified histoarchitecture in bladder tumor that arises from the degradation of the extracellular matrix by malignant cells during carcinogenesis may affect the penetration and uptake of hypericin by altering the efficiency of paracellular diffusion of the intravesically instilled hypericin for different grades of bladder cancers. Furthermore, it is also known that the expression of such transmembrane glycoproteins as E-cadherin decreases with decreasing level of differentiation in cell lines.³⁰

In relation to our observed results, we propose that the relatively well conserved cellular matrix and tight cellular adhesion due to high level of E-cadherin expression in well-differentiated low-grade bladder cancers may provide barriers to paracellular diffusion of hypericin, which reduces its inward diffusion efficiency during the period of instillation. As the hypericin is taken up by the urothelial cells, their poor diffusion efficiency tends to favor their accumulation at the superficial urothelial region.¹⁸ This urothelial accumulation is also supported in previous studies by Kamuhabwa, who observed that despite increasing the instillation times and concentration gradient of hypericin at the surface, the extent of hypericin penetration across the bladder wall was not significantly affected.^{31, 32} For poorly differentiated high-grade bladder cancers, the largely degraded cellular matrix due to loss of E-cadherin expression facilitates paracellular diffusion of hypericin from the urothelium toward the tumor interior. For this reason, the hypericin taken up by the high-grade urothelial cells has a higher tendency to reach the tumor interior, with a smaller accumulation at the superficial urothelial legion.

Furthermore, as the $440\text{-}\mathrm{nm}$ blue excitation light used in this study can penetrate tissue up to a depth of approximately $250\mu \mathrm{m}$ ,³³ while a typical three to five cell layers thick normal epithelium is approximately $50\mu \mathrm{m}$ ,³⁴ this excitation light may be able to excite only the hypericin localized in the superficial urothelium or at most into the submucosa layer ( $300\text{-}\mu \mathrm{m}$ thick³⁵) in normal bladder. For transitional cell carcinoma in which the urothelium thickens to more than 10 cell layers $(>250\mu \mathrm{m})$ ,³⁴ the penetration of the excitation light may be further confined to a limited extent in the urothelium, hence limiting the amount of hypericin that can be excited to the superficial urothelium layer. Hence, the greater accumulation of excitable hypericin in the urothelium of lower-grade bladder cancers due to impeded paracellular transport suggests a possible explanation for their greater fluorescence intensity ratios compared to higher-grade bladder cancers. While the preceding discussion may provide a possible account for our observation, it is also worth noting that such a hypothesis is limited and speculative in nature, and a separate study based on clinical data on glycoprotein expressions and hypericin distribution in clinical biopsies would otherwise provide more valuable insights to our macroscopic observations.

3.4.

Performance of Linear Discriminatory Function in Grading

The performance of the diagnostic intensity ratios in discriminating between the different grades of malignant tissues were subsequently assessed to determine their sensitivity and specificity in analyzing the grade of bladder cancer in situ. Each of the cystoscopic images was mapped onto a two-dimensional (2-D) scatter map based on both its $I_R∕I_G$ and its $I_R∕I_B$ values, and linear discriminatory functions were developed to discriminate the scatter map into different regions for classification of the lesions in the cystoscopic images into different grades of bladder cancer. Figure 5 shows the 2-D scatter plot of the intensity ratio $I_R∕I_G$ with respect to $I_R∕I_B$ for all the image data points in this study. Using the developed linear discrimination function that intersected the $I_R∕I_G$ axis at 3.66 and $I_R∕I_B$ at 1.06, i.e., $I_R∕I_G=-3.46I_R∕I_B+3.66$ , the combined two diagnostic intensity ratios yielded a sensitivity and specificity of 100.0% and 63.6%, respectively, for differentiating malignant from benign tissue. Using the other two developed linear discrimination functions with equations $I_R∕I_G=-3.46I_R∕I_B+5.88$ and $I_R∕I_G=-3.46I_R∕I_B+11.44$ , as shown in Fig. 5, we were able to further separate the malignant tissue, i.e., grade III from grade I/II and grade I from grade II/III, respectively. When combined together, these two linear functions were able to separate all three different grades of malignant bladder cancer simultaneously with mean sensitivity and specificity of 68.6% and 86.1%, respectively. The exact level of sensitivity and specificity achieved for classifying each of the different grades of malignant tissue based on all three linear discrimination functions is shown in Table 3 . The results obtained from these linear decision lines show that the combination of the two diagnostic intensity ratios can be used to generate possible discriminating criteria that achieve a promising level of sensitivity and specificity grading of bladder cancer in situ.

Fig. 5

Two-dimensional scatter map of $I_R∕I_G$ versus $I_R∕I_B$ for all the image data points in this study. The three diagnostic linear discriminatory functions are also shown, which discriminate the map into four different regions of benign condition and grade I, grade II, and grade III TCC.

Table 3

Level of sensitivity and specificity achieved in grading the bladder cancer using the developed linear discriminatory functions in the 2-D scatter plot of IR∕IG versus IR∕IB .

While we have provided possible accounts for our observation on the fluorescence correlation with cancer grade, a clearer understanding of the mechanisms leading to the observed trend warrants a more detailed study. However, for the purpose of this study, we have demonstrated the clinical potential of applying image analysis on hypericin fluorescence cystoscopic images not just to detect malignant lesions from benign surroundings, but also to assist clinicians in further classifying the malignant lesions into different pathological grades in situ and in real time based on just the cystoscopic images. Our reported findings based on the application of the three linear discrimination functions on our data set showed that this image analysis technique can produce promising performance on the sensitivity and specificity of detecting each grade of bladder cancer. It is noted from our results that the specificity in classifying each of the three grades of bladder cancer is generally higher compared to the sensitivity in picking up each of the different grades. This is in contrast to the performance achieved in discriminating malignant lesions from benign surroundings where the sensitivity is higher compared to the specificity (Table 3). While a highly sensitive technique may be more desirable for detecting malignant lesions, such a technique with high specificity for classification of different cancer grades will be desirable to ensure an accurate classification of the cancers. Furthermore, it is noted that the specificity in detecting grade II TCC is slightly lower than grade I and grade III TCC, and this may be attributed to its mid-point classification, which makes this category of pathology susceptible to false positives from neighboring categories at both ends. Also, this technique seems to be increasingly more sensitive in detecting higher-grade bladder cancers. As the sample size involved in this study is modest, we note that the results obtained are only approximate indications to the performance of this technique and the actual performance warrants a larger-scale study involving more image data points. A larger study coupled with more complex and refined discrimination functions derived from a larger set of image data points could further fine-tune and improve the sensitivity and specificity performance of this technique.