Idea Transcript
0016-5107/95/4106-057753.00 + 0 GASTROINTESTINALENDOSCOPY Copyright © 1995 by the American Society for Gastrointestinal Endoscopy
Spectroscopic diagnosis of esophageal cancer: new classification model, improved measurement system Masoud Panjehpour, PhD, Bergein F. Overholt, MD, James L. Schmidhammer, PhD Christie Farris, BS, Paul F. Buckley, BS, Tuan Vo-Dinh, PhD Oak Ridge and Knoxvi//e, Tennessee
Laser-induced fluorescence spectroscopy was used to measure fluorescence emission of normal and malignant tissue during endoscopy in patients with esophageal cancer and volunteers with normal esophagus. The spectroscopy system consisted of a nitrogen-pumped dye-laser tuned at 410 nm for excitation source, an optical multichannel analyzer for spectrum analysis, and a fiberoptic probe designed for both the delivery of excitation light and the collection of fluorescence emission from tissue. The fluorescence lineshape of each spectrum was determined and sampled at 15-nm intervals from 430 to 716 nm. A calibration set of spectra from normal and malignant spectra was selected. Using stepwise discriminate analysis, significant wavelengths that separated normal from malignant spectra were selected. The intensities at these wavelengths were used to formulate a classification model using linear discriminate analysis. The model was then used to classify additional tissue spectra from 26 malignant and 108 normal sites into either normal or malignant spectra. A sensitivity of 100% and specificity of 98% were obtained. (Gastrointest Endosc 1995;41:577-81.) Laser-induced fluorescence of tissue has been used to distinguish normal from malignant and premalignant tissues. 1-s Several measurement systems and data analysis techniques have been employed to separate normal from malignant tissue spectra, with variable degrees of success. Typically, laser light is used for excitation of the tissue. Fluorescence emission is collected by a fiberoptic probe and delivered to a spectrum analyzer. The fluorescence emission is analyzed using a spectral analysis technique. An empirical 1, 2, 4 or statistical 3, 5 model based on the spectral Received January 10, 1994. For revision April 29, 1994. Accepted July 25, 1994. From the Laser Center, Thompson Cancer Survival Center, Knoxville; Oak Ridge National Laboratory, Oak Ridge; and Departrnent of Statistics, University of Tennessee, Knoxville, Tennessee. This work was supported by funds from Thompson Cancer Survival Center, Thompson Charitable Foundation, and American Laser Foundation in Knoxville, Tennessee. Reprint requests: Masoud Panjehpour, PhD, Laser Center, Thompson Cancer Survival Center, 1915 White Avenue, Knoxville, Tennessee 37916. 37/1/60118 VOLUME 41, NO. 6, 1995
analysis is commonly used to create a classification criterion to differentiate normal from malignant tissue spectra. The accuracy and clinical acceptance of this technique may be enhanced by making additional refinements, such as (!) determining an excitation wavelength for each organ system that classifies normal and malignant tissues in vivo with a high degree of accuracy, (2) developing a diagnostic classification model suitable for observations with a nonnumeric dependent variable (normal or cancer), and (3) assembling a user-friendly system for routine clinical use. In this study, we (1) tested a new excitation wavelength for diagnosis of esophageal cancer, (2) developed a new classification model using linear discriminate analysis to diagnose esophageal cancer with a high degree of accuracy, and (3) employed a physicianfriendly system that allows rapid and reliable in vivo fluorescence measurements. MATERIALS AND METHODS The study was approved by the Institutional Review Board of the T h o m p s o n Cancer Survival Center. Before any
GASTROINTESTINAL ENDOSCOPY 577
measurements, informed consent was obtained. A total of 32 patients were entered into the study. Laser-induced fluorescence measurement system A nitrogen-pumped dye-laser (model LC300C, Laser Photonics, Inc., Orlando, Fla.) was used to deliver 5-ns pulses of excitation light. The dye-laser was tuned at 410 nm. The maximum repetition rate was 40 Hz. The pulse energy was highly stable, typically at 20.0 ___0.3 #J. The laser was externally activated by an optical multichannel analyzer system. The laser beam was coupled via a quartz lens to the excitation fibers of a sheathed fluorescence probe (C Technologies, Verona, N.J.). This probe consisted of seven 200-gm fibers for delivery of the excitation light and twelve 200-ttm fibers surrounding the excitation fibers for collection of tissue fluorescence emission. The distal end of the probe was flat and encased in a 10-mm-long tube with outside diameter of 1.7 mm. The output (proximal) end of the collection fibers was arranged in a linear array. The length of this fiber assembly was 6 m, allowing easy use during endoscopy procedures. The overall outside diameter of the probe was 2.4 mm (including the protective sheath). The linear array of collection fibers formed the input to the entrance slit of an f/3.8, 0.28-m, triple-grating spectrograph (mode11235, EG&G Princeton Applied Research, Princeton, N.J.). A grating of 150 grooves per millimeter was selected to allow wide-band analysis of the fluorescence spectra. The center wavelength of the spectrograph was set at 733 nm in order to shift the excitation component of the collected light 20 nm oft the detector. The spectrally dispersed emission spectra were imaged on a gated intensified 1024-diode array detector (model 1456B990-HQ) controlled by an optical multichannel analyzer (OMA III, EG&G Princeton Applied Research). Synchronization of laser pulses and gating of the intensified diode array were achieved by programming a 1336-ns delay between laser trigger and detector activation. The intensifier was gated for 100 ns, during which a 5-ns laser pulse was delivered to the tissue. The diode array was spectrally calibrated using a mercury-argon spectral calibration lamp (model 6035, Orie! Corp., Stratford, Conn.). A personal computer (Gateway 2000, model 486DX2, 50 MHz) was used to control the entire system. OMA-VisionPDA spectroscopy software (EG&G Princeton Applied Research) was utilized to conduct the measurements. This software allowed automatic storage of each spectrum into the hard disk with a filename coded using a combination of numbers representing the current date and a three-digit file number. Each measurement was conducted and stored on the hard disk by depressing a single key. A foot switch was attached to the keyboard to allow remote activation of the measurements. Typically, the assistant set up the system before each procedure. The physician then controlled all operations simply by pressing the foot switch, which required less than a second. Each measurement required i to 2 seconds for processing and storage. The software was operated in accumulate/subtract mode. In this mode, three operations form a measurement cycle: (1) The OMA system sends a trigger signal to the excitation laser. (2) After a programmed delay of 1336 ns, the OMA system also sends a signal to the detector to be activated. (3) The 578
GASTROINTESTINAL ENDOSCOPY
OMA system performs background measurement, subtracts the background from the signal, and stores/adds the result in memory. In this study, 10 emission scans from tissue were accumulated and the result was stored in the hard disk. Overall system response In order to assure optimal system performance, the fiuorescence of a standard dye solution (Kiton red) and the laser pulse energy were measured between all patients. In this manner, any variation in fluorescence signal caused by laser energy, detector gain, or probe damage could be detected and corrected immediately. It should be emphasized that fluorescence lineshape is independent of fluorescence emission intensity. Therefore, laser pulse energy and standard dye fluorescence intensity were not needed for calcutation of fluorescence lineshape. Endoscopic fluorescence measurement Fluorescence measurements were performed during routine UGI endoscopy procedures. The endoscopist passed the fluorescence probe through the biopsy channel of the endoscope and touched it lightly to the tissue. Measurements were initiated by pressing the foot switch. Typically, three fluorescence measurements were recorded from each site. The collected spectra were automatically saved in separate coded data files. Fluorescence measurements were conducted from areas of normal esophageal mucosa and sites of esophageal cancer. A total of 159 fluorescence measurements from 32 patients were used in this study: 123 spectra from normal mucosa and 36 from esophageal carcinoma. Three of the spectra from malignant tissue were obtained from squamous cell carcinoma and the rest from adenocarcinoma arising from Barrett's mucosa. Fluorescence measurements of normal esophageal mucosa were obtained from both cancer patients and other volunteers who had otherwise normal esophagi. In cancer patients, normal-appearing mucosa was selected 10 to 15 cm proximal to the cancer site. In patients with Barrett's syndrome, readings of normal tissue were obtained from an area 10 to 15 cm proximal to the squamocolumnar junction. Great care was taken to assure that measurements were taken from viable malignant tissue. In fact, proper placement of the probe against the tissue was verified on the endoscopy monitor and noted during each measurement. If verification was not possible, the measured fluorescence was discarded. Although some measurements were conducted from areas of Barrett's mucosa, those data were not included in this study. A biopsy specimen was taken from each abnormal lesion for pathology studies. Initially, biopsy specimens ware also taken from normal mucosa, but this was found unnecessary and was discontinued. The results of pathology studies were used to classify the fluorescence data into either a normal or malignant category. The biopsy results were used as the gold standard to evaluate the reliability of this diagnostic technique. Formulation of the classification algorithm Fifteen spectra from normal mucosa and 10 spectra from malignant tissue were used as a calibration data set to determine a classification model. The lineshape of each fluV O L U M E 41, NO. 6, 1995
orescence spectrum was determined by dividing the spectrum by the area under the spectrum from 430 to 716 nm (total amount of light from 430 to 716 nm). Each fluorescence lineshape spectrum was sampled at 15-nm intervals from 430 to 716 nm, excluding 430 nm. This resulted in intensities at 19 wavelengths being used in the statistical analysis. Becanse tissue fluorescence emissions have a broad-band spectral shape without any sharp features, 15-nm sampling represented the spectra without any loss of information5 (Nyquist criterion). The intensities at the sampled wavelengths were used as the independent variables in a stepwise discriminate analysis. The dependent variable had two nonnumeric values, N for normal and C for cancer. The « to enter and remove variables from the model was 0.15. Using the stepwise discriminate analysis, the significant wavelengths (optimal set) were selected. The intensities at the selected wavelengths were used in formulating a classification model using linear discriminate analysis. After determination of the classification model, the fluorescence lineshape from 108 normal and 26 malignant tissue specimens were used to determine the sensitivity and specificity of the model in classifying malignant and normal tissues. The model calculated a probability of membership for each spectrum classification. The probability of membership represents the classification confidence level.
RESULTS Laser-induced fluorescence spectra of 123 sites of normal esophageal mucosa and 36 sites of carcinoma were recorded during routine endoscopy. The lineshape of each spectrum was determined. It was first verified that no significant difference existed between the fluorescence emission of normal mucosa in cancer patients and that of patients without cancer. Figure 1 shows typical fluorescence lineshape spectra obtained from normal mucosa and carcinoma of the esophagus. The fluorescence lineshapes are clearly different in normal and malignant tissue. The two spectra intersect at about 540 nm. Below 540 nm, the normal tissue fluorescence is stronger than that of malignant tissue. Above 540 nm, the normal tissue fluorescence is weaker than that of malignant tissue. Stepwise discriminate analysis was used to derermine the minimum number of sampled wavelengths whose values would be used to form a discriminate function. The fluorescence intensities at 19 sampled wavelengths were entered into the analysis. The stepwise discriminate analysis selected only five wavelengths: 490 nm, 580 nm, 670 nm, 685 nm, and 715 nm. An average squared canonical correlation (correlation coefficient) of 0.96 was obtained, indicating a high degree of confidence in using the above wavelengths (and the 25 fluorescence spectra from normal and malignant tissue used as the calibration set) to construct a model to differentiate normal from malignant spectra. The correlation coefficient was improved from 0.90 (using the intensity at only one wavelength) to 0.96 (using the V O L U M E 41, NO. 6, 1995
0.006-
0.005-
0.004-
2~ G)
0.003-
0.002-
0.001 -
0 430
4~0 4~0 5~0 s~o 5~0 6~0 64o 8~0 7öo
730
Wavelength (nm)
Figure 1. Typical fluorescence lineshapes of normal mucosa (solid line) and carcinoma (dotted line) of the esophagus. intensities at five wavelengths). The correlation coefficient was not improved by entering additional wavelengths into the model. The correlation coefficient value could be improved by adding more fluorescence spectra to the calibration set. In this study, correlation of 0.96 was satisfactory. The intensities at these wavelengths were used to create a classification model using linear discriminate analysis. The accuracy of the model in classifying different spectra was tested by entering 134 spectra (108 normal and 26 malignant) into the model. The model automatically classified each spectrum into either a normal or malignant category with a corresponding probability of membership. Overall, 132 of 134 spectra were correctly classified. Two of 108 spectra from normal tissue were classified as malignant; the other 106 were classified correctly. All 26 spectra from malignant tissues were classified correctly. This corresponds to a sensitivity and specificity of 100 % and 98 %, respectively. In addition, in 129 of the correctly classified spectra, the probability of membership into the classified group was 100 %, indicating the highest degree of confidence. Three normal tissue spectra were classified correctly with probabilities of 99.75%, 99.88%, and 99.89%.
DlSCUSSlON Laser-induced fluorescence spectroscopy is a noninvasive technique that may be used to differentiate normal from malignant tissue. This technique capitalizes on the principle that when certain compounds are excited by light, they exhibit a characteristic fluorescence emission. Laser-induced fluorescence in combination with photosensitizing drugs has been used to detect malignant tissues. 6 Dihematoporphyrin ether (Photofrin) is a photosensitizer used for both diagnoGASTROINTESTINAL ENDOSCOPY
579
sis and photodynamic therapy of cancer. Profio and Balchum 7 provided a comprehensive review of this diagnostic technique. However, a disadvantage associated with the use of photosensitizers is that skin photosensitivity lasting 4 to 6 weeks is induced. It is desirable to develop a fluorescence technique that allows differentiation between normal and malignant tissue without the need for an exogenous fluorescent agent. Recently, several reports have suggested the presence of endogenous fluorescence differences between normal and malignant tissue. Alfano et al. 1 used different laser lines from an argon laser in vitro to excite fluorescence of normal and malignant tissues from lung and breast. Using two pairs of malignant and normal breast and lung tissue, they showed marked differences between spectra from normal and cancerous samples. Additional data from this study have been published by Tang et al.2 Fluorescence imaging has been used to detect dysplasia and carcinoma in situ in lung. Lam et al.s described a fluorescence bronchoscope used to complement white-light bronchoscopy. They stated that the sensitivity of fluorescence bronchoscopy was 50 % greater than that of white-light bronchoscopy in detecting dysplasia and carcinoma in situ. Laser-induced fluorescence has been used for detection of premalignant lesions of the gastrointestinal tract. In an in vitro study, Kapadia et al.5 used 325-nm light from a helium-cadmium laser for differentiation of adenomatous polyps from normal mucosa and hyperplastic polyps. Using stepwise multivariate linear regression analysis, they developed a classification algorithm. They indicated that all 34 normal mucosal specimens and all 16 adenomatous polyps were classified correctly. Of 16 hyperplastic polyps, 15 were classified as normal tissue. Schomacker et al.3 used a nitrogen laser (337 nm) in vivo and in vitro to excite fluorescence of colonic tissue. Using a multivariate linear regression analysis, they distinguished neoplastic tissue from tissue that was not neoplastic with sensitivity and specificity of 80 % and 92 %, respectively. Sensitivity and specificity of 86 % and 77 % were obtained when multivariate linear regression analysis was used to differentiate neoplastic polyps from polyps that were not neoplastie. In a similar study, Sehomacker et al. 9 studied the changes in tissue fluorescence occurring postmortem. They indicated that in vivo and in vitro fluorescence emissions were significantly different in adenomatous polyps and concluded that laser-indueed fluorescence data should be acquired in vivo. In an in vitro study, Richards-Kortum et als ° used fluorescence excitation-emission matrices to identify the optimal excitation wavelength for obtaining fluorescence emission spectra that could be used to differentiate normal from pathologic colonic tissue. They 580
GASTROINTESTINAL ENDOSCOPY
indicated that excitation at 330, 370, and 430 _+ 10 nm was optimal for this purpose. Cothren et al. 4 used 370-nm excitation light to examine in vivo fluorescence of colonic tissue. Using a two-dimensional scatter plot of fluorescence intensities at 460 nm versus 680 nm, they defined a straight line representing a decision surface, which minimized the number of misclassified samples. Using this technique, adenomas could be distinguished from nonadenomatous tissue in approximately 97 % of cases. The literature clearly shows the feasibility of laserinduced fluorescence technique to diagnose cancer. However, several improvements can enhance the sensitivity and clinical acceptance of this powerful diagnostic technique. Important factors in using laser-induced fluorescence spectroscopy to diagnose cancer are ease, reliability, and reproducibility of the measurements. Often, it is difficult to measure fiuorescence of gastrointestinal tissue endoscopically because of motility, respiratory movements, and cardiac activity. The reliability of the measurements can be enhanced by improving the overall system to conduct the measurements rapidly. In the system described here, a 50-MHz computer with a math coprocessor was used, allowing operation of the laser-induced fluorescence system at a high-capacity rate. In addition, the spectral analysis software automatically stored each spectrum into the hard disk without the need for any operator interaction. Allowing the endoscopist to use a foot switch to initiate the measurements permitted visual/manual synchronization of measurement and endoscopic placement of the probe against tissue. Typically, 20 fluorescence measurements could be conducted in less than I minute. Rapid and reliable measurements proved invaluable in a clinical environment. In this study, the excitation wavelength of 410 nm proved to be suitable for diagnosis of esophageal malignancies. This is in close agreement with data from Richard-Kortum et al. 1° reporting excitation wavelength of 430 + 10 nm for classification of colonic lesions. Using a 405-nm line from a krypton laser, Hung et al.ll showed significant spectral differences between tiuorescence of normal and malignant bronchial tissue. The 410-nm excitation wavelength was tested because it was close to the excitation wavelengths reported elsewhere,lO, 11 it had not been evaluated for esophageal cancer, and our pilot study proved it suitable for this indication. In addition, the 410-nm wavelength could be used to detect fluorescence of dihematoporphyrin ether in tissue, allowing application of this system in multiple projects. However, although 410-nm excitation was suitable for this study, it does not mean it is the optimal wavelength. This excitation wavelength may or may not be appropriate for diagnosis of other malignancies. Each excitation wavelength should V O L U M E 41, NO. 6, 1995
be t e s t e d extensively to d e t e r m i n e its suitability for the diagnosis of a given malignancy. An i m p o r t a n t factor in developing a reliable diagnostic t e c h n i q u e is the use of a p p r o p r i a t e statistical analysis. Although multivariate linear regression analysis has been used to develop classification algorithms, it is not i n t e n d e d for cases in which the d e p e n d e n t variable is n o n n u m e r i c in n a t u r e (normal or malignant) and the i n d e p e n d e n t variable(s) has continuous quantitative values (fluorescence intensity at different wavelengths). In this case, it is more appropriate to use linear discriminate analysis to d e t e r m i n e a classification model based on a calibration set of spectra from normal and malignant tissues. Using this model, subsequent spectra can be classified into either a normal or a malignant group with a corresponding probability of m e m b e r s h i p in t h a t group. This is unlike other techniques, in which a classification threshold is selected arbitrarily 4 or a quantitative score is f o r m u l a t e d to classify normal spectra with a positive score and malignant spectra with a negative score. 3, 5, 9 T h e use of the fluorescence lineshape improved the sensitivity of this diagnostic technique. Fluorescence lineshape represents the distribution of different wavelengths within the integrated fluorescence emission. It also represents the chemical composition of the tissue c o m p o u n d s (in terms of fluorescence emission, absorption, or both). In addition, because the fluorescence lineshape is i n d e p e n d e n t of the fluorescence intensity, it eliminates the effect of system p a r a m e t e r s such as laser energy, detector gain, and optics efficiency. Although these p a r a m e t e r s were not needed for lineshape calculations, t h e y were m o n i t o r e d to assure optimal system performance. We are c u r r e n t l y investigating the use of laserinduced fluorescence spectroscopy to detect dysplasia and early cancer in patients with B a r r e t t ' s esophagus. Our preliminary d a t a indicate t h a t fluorescence of normal B a r r e t t ' s mucosa in a p a t i e n t with B a r r e t t ' s esophagus can be distinguished from t h a t of severe dysplasia in patient's with B a r r e t t ' s esophagus. In summary, a clinically user-friendly laser-induced fluorescence system has been described hefe to measure in vivo fluorescence of gastrointestinal tissue during routine endoscopy. A classification model, spe-
V O L U M E 41, NO. 6, 1995
cifically designed for binary non-numeric observations, was developed t h a t automatically classified each spectrum into either a normal or a malignant group accurately and with a high probability of membership. Finally, this technique has the potential to provide real-time diagnosis of malignant and premalignant tissue in the esophagus. Diagnosis and t r e a t m e n t of cancer m a y be conducted during the same endoscopy procedure, eliminating unnecessary delay and rep e a t e d procedures. Because this technique is noninvasive, an unlimited n u m b e r of sites on the surface of the lesion m a y be examined to improve early detection of malignancies. In addition, a fluorescence gastrointestinal endoscope m a y be developed to improve the sensitivity of white-light endoscopy in detecting malign a n t and premalignant tissue, s REFERENCES
1. AlfanoRR, Tang GC, Pradhan A, Lam W, Choy DSJ, Opher E. Fluorescence spectra from cancerous and normal human breast and lung tissues. IEEE Journal of Quantum Electronics 1987; 23:1806-11. 2. Tang GC, Pradhan A, Alfano RR. Spectroscopic differences between human cancer and normal lung and breast tissues. Lasers Surg Med 1989;9:290-5. 3. Schomacker KT, Frisoli JK, Compton CC, et al. Ultraviolet laser-induced fluorescenceof colonic tissue: basic biology and diagnostic potential. Lasers Surg Med 1992;12:63~78. 4. Cothren RM, Richards-Kortum R, Sivak MV Jr, et al. Gastrointestinal tissue diagnosis by laser-induced fluorescence spectroscopy at endoscopy. Gastrointest Endosc 1990;36:10511. 5. Kapadia CR, Cutruzzola FW, O'Brien KM, Stetz ML, Enriquez R, Deckelbaum LI. Laser-induced fluorescencespectroscopy of human colonicmucosa--detection of adenomatous transformation. Gastroenterology 1990;99:150-7. 6. Dougherty TJ. Photosensitizers: therapy and detection of malignant tumors. Photochem Photobiol 1987;45:879-89. 7. Profio AE, Balchum OS. Fluorescence diagnosis of cancer. In: Kessel D, ed. Methods in porphyrin photosensitization. New York: Plenum Press, 1985:43-50. 8. Lam S, MacAulay C, Hung J, LeRiche J, Profio AE, Palcic B. Detection of dysplasia and carcinoma in situ with a lung imaging fluorescence endoscope device. J Thorac Cardiovasc Surg 1993;105:1035-40. 9. Schomacker KT, Frisoli JK, Compton CC, et al. Ultraviolet laser-induced fluorescence of colonic polyps. Gastroenterology 1992;102:1155-60. 10. Richards-Kortum R, Rava RP, Petras RE, Fitzmaurice M, Sivak M, Feld MS. Spectroscopic diagnosis of colonic dysplasia. Photochem Photobiol 1991;53:777-86. 11. Hung J, Lam S, LeRiche J, Palcic B. Autofluorescence of normal and malignant bronchial tissue. Lasers Surg Med 1991;11: 99-105.
GASTROINTESTINAL ENDOSCOPY
581