We report a multimodal optical approach using both Raman spectroscopy and optical coherence tomography (OCT) in tandem to discriminate between colonic adenocarcinoma and normal colon. Although both of these non-invasive techniques are capable of discriminating between normal and tumour tissues, they are unable individually to provide both the high specificity and high sensitivity required for disease diagnosis. We combine the chemical information derived from Raman spectroscopy with the texture parameters extracted from OCT images. The sensitivity obtained using Raman spectroscopy and OCT individually was 89% and 78% respectively and the specificity was 77% and 74% respectively. Combining the information derived using the two techniques increased both sensitivity and specificity to 94% demonstrating that combining complementary optical information enhances diagnostic accuracy. These data demonstrate that multimodal optical analysis has the potential to achieve accurate non-invasive cancer diagnosis.
© 2013 Optical Society of America
Colorectal adenocarcinoma is the fourth commonest cancer in the UK  and the US . The investigation of patients with suspected colorectal cancer, or the surveillance of individuals at high risk of the development of colorectal cancer, is currently performed by endoscopic assessment of the colorectal mucosa using colonoscopy. This requires visual assessment of the mucosa, with targeted biopsy of any abnormal areas. Both the colonoscopic assessment, and the subsequent histopathological assessment of the biopsies, is subjective. The development of novel methodologies that enhance the endoscopist’s ability to assess the colorectal mucosa would therefore not only improve the targeting of biopsies to the most appropriate areas, but would also potentially obviate the need for a tissue biopsy.
The biochemical features associated with normal and neoplastic tissues are distinct, reflecting differences in the biological processes present . However, the change in the concentration of these biomarkers is relatively small and thus very sensitive and powerful techniques are required for their detection. Clinical studies on colon cancer and pre-cancer tissues using fluorescence spectroscopy have shown that the spectra are very similar to benign abnormalities in some patients and hence hard to classify . Near-IR Raman spectroscopy studies on colon cancer tissue demonstrated the efficiency of Raman spectroscopy in identifying neoplastic and normal tissues [5, 6]. However, although some previous studies show high sensitivity and specificity in discriminating between normal and adenocarcinoma colonic tissues, the Raman spectra illustrated in these studies show that the Raman fingerprints of normal and abnormal tissues are very similar. It has also been emphasized that inter-patient variability is present in the chemical composition of colonic tissues . This means that, when considering inter-patient variability , the classification efficiency may be below acceptable limits for using this as a reliable tool for disease diagnosis.
One solution to this issue would be to take a multimodal approach and combine Raman spectroscopy with other methods that provide complementary information about the tissue under investigation. Optical coherence tomography (OCT) is a powerful, non-invasive optical imaging technique that relies on scattering from the sample and subsequent interference, and is capable of providing cross-sectional images of tissues [8, 9]: morphological information can be obtained with an axial resolution ~6.2 µm. While visual inspection of OCT images may be used to assess tissue morphology, it has been demonstrated that image processing approaches can achieve automated extraction of features from OCT images and predict the state of tissues. In particular, texture analysis of OCT images has been successful in discriminating between normal and abnormal tissue types [10, 11].
Previous studies have indicated that a combination of Raman spectroscopy and OCT may be used as an effective multimodal tool for disease diagnosis [12–14]. There are two challenges related to the implementation of a combined Raman-OCT system – integration of the hardware and analysis of the complementary information obtained from both modalities. The previous studies mainly focused on the former issue [12, 14].
In this study we address the latter issue and demonstrate how the information from these complementary modalities, which are fundamentally different in representation, may be combined to enhance the sensitivity and specificity of this multimodal diagnostic tool. We quantitatively combined the biochemical information from Raman spectroscopy and the morphological information from OCT. We assessed the accuracy of the classifier using individual information as well as multimodal information and observed that there is a clear increase in the accuracy of the classifier when multimodal information is combined. This demonstrates the potential of developing multimodal systems by combining Raman spectroscopy and OCT, which in combination perform an efficient non-invasive optical biopsy.
2. Materials and methods
2.1 Tissue samples
All the tissue samples used in this study were obtained from the Tayside Tissue Bank, Ninewells Hospital and Medical School, Dundee (Tissue request no. TR000289) with appropriate ethical permission. The tissue samples were snap-frozen a few minutes after dissection and stored at −80 °C. Each sample had a dimension of approximately 2mm x 2mm x 1mm.
The tissue samples used comprised 33 normal tissues and 18 paired tumour tissues. 11 further samples where both normal and tumour tissue were present in the sample (referred hereafter as “mix”) were also analyzed. For optical interrogation, the tissue sample was thawed for 5 minutes and placed on a quartz coverslip (22 mm x 22 mm, thickness 180 µm) (SPi supplies). From each tissue sample, one OCT image and one Raman spectrum were acquired. Due to the collection geometry of the optical methods used in this study, the multimodal data acquired from the tissue sample were not spatially co-registered. After image and spectral acquisition, the sample was fixed in 10% neutral buffered formalin solution for histological evaluation.
2.2 Raman spectroscopy
The Raman spectra from tissues were obtained using a confocal Raman spectroscopic system built around an inverted biological microscope (Nikon, Eclipse TE2000). The system was equipped with a diode laser (Sacher Lasertechnik, maximum power 1 W) for Raman excitation. For detection, the system was equipped with a monochromator (Shamrock SR-303i, Andor Technology) with a 400 lines/mm grating, blazed at 850 nm and a deep depletion, back illuminated and thermoelectrically cooled CCD camera (Newton, Andor Technology). The laser beam was focused through a 40x objective (Nikon, NA 0.65) onto the sample delivering a power of 50mW at the sample plane. With a confocal aperture of diameter 500 µm, the confocal cylinder from which the signal was collected had a dimension of base diameter 12.5 µm and height 10 µm. The power used was below the power level required for disruption to the morphological or chemical composition of the tissue .
Tissue samples, placed on a quartz coverslip, were loaded onto the confocal Raman microscope. The back-reflection of the excitation beam was used to ensure that the confocal volume was within the sample and the position of the sample volume from the quartz coverslip was the same for all samples. The Raman spectrum from each sample was acquired with an acquisition time of 5 s.
2.3 Optical coherence tomography
The OCT images were acquired utilizing a home-built Fourier domain optical coherence tomography (FDOCT) system . FDOCT allows the acquisition of a cross-sectional image of the sample of interest avoiding scanning of the reference arm through the depth range [17, 18]: the back-reflected or back-scattered signal from the sample is acquired in a single event and the Fourier transform of the collected spectrum delivers the depth information from the sample.
The FDOCT system used in this study consisted of a fiber-coupled superluminescent diode (SLED371-HP1, Superlum Diodes) as a short-coherence light source with a bandwidth of 50 nm centred at 840 nm. The sample arm directed light onto the tissue by a telecentric microscope objective with a working distance of 25 mm (LSM03-BB, Thorlabs). The signal was detected by a custom spectrometer formed by a dispersive optical element (volume phase holographic grating, 1200 l/mm, 830 nm, Wasatch Photonics) and a line CCD detector (Aviiva EM1, 2048 pixels, pixel size 14 × 28 μm, e2v). Using a high-reflectivity mirror (PF10-03-M01, Thorlabs) as a test sample and a neutral density filter (ND40B, Thorlabs) along the sample arm, the signal-to-noise ratio (SNR) of the system has been measured to be 97 dB with a CCD integration time of 80 μs. The SNR may be increased by increasing the integration time .
The FDOCT system had an axial resolution and a lateral resolution of approximately 6.2 μm and 17 μm respectively. The full depth range was 1.7 mm in air. Considering a typical colon refractive index of 1.4 , the maximum depth range decreases to about 1.2 mm. The lateral scan range can be set up to 5 mm with the galvanometric mirror with 512 lines acquired. In particular, the field of view of the images presented in this work was restricted to about 1.1 mm in depth and 3.2 mm wide. The total acquisition time for a full image was 1 s as a consequence of CCD integration time, data processing, and time to display on screen.
Each tissue sample was mounted on a quartz slide prior to acquisition of cross sectional images from the OCT system.
After the acquisition of Raman and OCT data, the tissue samples were fixed in neutral buffered formalin and processed to paraffin wax. 5µm sections were cut and stained with haematoxylin and eosin for histological evaluation. The histological assessment was performed by an experienced clinical pathologist (CSH) and representative histological images from various categories of tissues used in this study are shown in Fig. 1 .
2.5 Raman spectra analysis
All the Raman spectra acquired from the tissue were smoothed with a Savitzky – Golay filter with smoothing width 9 and degree 3 followed by baselining using iterative modified polynomial fitting . The Raman bands that showed significant differences were identified using student’s t-test with a significance level of p<0.001. A comparison of Raman signatures from two types of tissue sample is shown in Fig. 2.
2.6 Texture analysis on OCT data
When combining a spectroscopic technique, which is usually analyzed quantitatively, with an imaging technique, which is usually analyzed qualitatively, it is essential to implement a quantification algorithm for the imaging technique. It has been demonstrated in previous studies that texture analysis is an efficient method to quantify OCT images [10, 11]. Texture is a measure of the variation of the intensity of a surface, quantifying properties such as smoothness, coarseness and regularity. Statistical techniques characterize texture by the statistical properties of the grey levels of the points comprising a surface. Typically, these properties are computed from the grey level histogram or grey level co‐occurrence matrix (GLCM) of the surface. 16 texture parameters were calculated from each image, namely “contrast”, “correlation”, “energy” and “homogeneity” in four directions (0°, 45°, 90°, 135°). The algorithm for evaluating the texture parameters and avoiding the non-tissue regions in the image are detailed in a previous study , which allowed calculation of the texture parameters independent of the morphology of the tissue surfaces. Typical OCT images from normal and tumour tissues are shown in Fig. 3.
2.7 Data treatment and discrimination of tissue types
Principal component analysis (PCA) was used for feature selection and reduction of parameters obtained for both Raman and OCT data. A supervised discrimination algorithm using a support vector machine (SVM) with a “linear” kernel was used to perform discrimination between the two tissue types. The sensitivity and specificity of each training data set was evaluated using “leave one out” cross validation (LOOCV). For the Raman and OCT data sets, the first 5 principal components (PC) were chosen to build the classifier. In the multimodal approach, the first 5 PCs of data from each modality were combined to build the classifier. A flow chart of the data analysis is given in Fig. 4.
3. Results and discussion
Figure 2 shows the spectral plot of the mean and the standard deviation of the Raman spectra collected from both the normal colonic mucosa and colonic adenocarcinoma from 850 cm−1 – 1750 cm−1. The prominent Raman bands corresponding to the amide I (1655), protein/lipid (1445), collagen/lipid (1320), amide III (1270), C-C stretch of lipids at 1085 and hydroxyproline (875) can be clearly seen in both cases [4, 7]. The standard deviation of the Raman spectra for both normal and tumour tissues is substantial and cannot be neglected. This may be attributed to inter-patient variability in the chemical composition of the colon tissues . The Raman bands that showed significant variations were identified by performing student’s t-test across the spectra with a significance level of p<0.001. The Raman bands that showed significant differences between normal tissue and colonic adenocarcinoma are: 855 cm−1 (ring breathing mode of Tyrosine), 931 cm−1 (C-C stretch of proline ring/protein), 1002 cm−1 (symmetric ring breathing mode of Phenylalanine), 1122 cm−1 (C-C stretch mode lipids/protein), 1180-1184 cm−1 (cytosine, guanine, adenine) and 1578 cm-1 (pyrimidine ring/nucleic acid) . Figure 5(a) shows the cluster plot of the first two PCs for the two tissue types along with the support vectors that discriminate the two tissue types. The sensitivity and specificity obtained for this data set were 89% and 77% respectively as shown in Table 1.
The representative OCT image shown in Fig. 3 demonstrates that there are some morphological differences between the two tissue types. However considering the inter-patient variability, visual examination of these images is not sufficient to reliably make the diagnosis. This is also evident from the sensitivity and specificity obtained from the texture analysis of OCT data, which were 78% and 74% respectively. Figure 5(b) shows the cluster plot of the first two principal components obtained from the texture parameters for the two tissue types along with the support vectors that discriminate between the two tissue types.
When multi-modal information was used in the analysis, combining the Raman spectroscopy and OCT data, the sensitivity and specificity were both increased to 94%. Figure 5(c) shows the cluster plot, where the first PCs of each modality were plotted against each other along with the support vectors that discriminate the two tissue types. Table 1 summarizes the classification results obtained for each modality.
Histological evaluation revealed that some tissue samples were a mix of both normal and tumour tissues. Figure 5(d) shows a clusterplot of the combined data set. The first PCs of each modality were plotted against each other for all three tissue groups. It can be seen that the “mix” sample clusters close to the intersection of normal and tumour samples, providing further validation of the method.
In this study there is a difference in the signal collection area for Raman and OCT. One area where the technique needs improvement is in achieving co-registration of the region from which the data is acquired. This study used only those tissues that were found to contain only normal or tumour tissue on histological evaluation for estimating the sensitivity and specificity. This ensured that lack of co-registration did not skew the classification results. Also the limited size of the tissue samples used prevented collection of data from multiple points of the same sample. While this study has demonstrated the potential of enhancing diagnostic accuracy using multimodal information, future research will be focused on developing suitable optical systems that can achieve spatial co-registration of the area from which this complementary information is collected.
Although previous studies have demonstrated very high sensitivity and specificity in tissue discrimination using Raman spectroscopy or OCT , our observations reveal that inter-patient variability may reduce test performance when implemented for practical applications. This is where multi-modal approaches that provide complementary information become relevant. The combination of OCT and Raman has been proposed as a potential multi-modal tool for disease diagnosis [12, 13].
When combining Raman spectroscopy and OCT, there are two major challenges. One is combining the hardware and the other is combining the information acquired from both of these modalities: we address the latter problem in this study. This is challenging especially considering the fact that representations of the information from these two modalities are different. Raman spectra need to be quantified to obtain useful inferences, while the OCT images are conventionally analyzed visually. The approach we put forward in this study to combine these two modalities involves quantifying OCT images using texture analysis. The texture information obtained from OCT may then be readily combined with the spectral information from Raman spectroscopy. Development of a suitable optical system that allows spatial co-registration of the region from which the morphological and spectral information are collected would help to translate such a multi-modal system for in vivo diagnosis.
When multi-modal information was used for tissue discrimination, the classifier resulted in an increase in both sensitivity and specificity (to 94%) compared to using single modalities. This shows the potential of combining Raman spectroscopy and OCT for implementing a reliable, optical diagnostic system. While colon tissue samples have been used here to demonstrate this potential, this approach has scope to be applied to the discrimination between normal and cancer tissues from other body sites.
We thank the UK EPSRC for funding, the CR-UK/EPSRC/MRC/DoH (England) imaging programme, the European Union project FAMOS (FP7 ICT, contract no. 317744) and the European Union project IIIOS (FP7/2007-2013, contract no. 238802). We thank Tayside Tissue Bank for providing us with the tissue samples under request number TR000289. K.D. is a Royal Society-Wolfson Merit Award Holder.
References and links
1. C. R. UK, “Bowel cancer Key Facts,” (Cancer Research UK, 2013), http://www.cancerresearchuk.org/cancer-info/cancerstats/keyfacts/bowel-cancer/#Bowel, Accessed 27/05/2013, 2013.
2. C. Scalfi-Happ, M. Udart, C. Hauser, and A. Rück, “Investigation of lipid bodies in a colon carcinoma cell line by confocal Raman microscopy,” Med. Laser Appl. 26(4), 152–157 (2011). [CrossRef]
5. M. S. Feld, R. Manoharan, J. Salenius, J. Orensteincarndona, T. J. Romer, J. F. Brennan, R. Dasari, and Y. Wang, “Detection and characterization of human tissue lesions with near infrared Raman spectroscopy ,” P Soc Photo-Opt Ins 2388, 99–104 (1995).
6. E. Widjaja, W. Zheng, and Z. W. Huang, “Classification of colonic tissues using near-infrared Raman spectroscopy and support vector machines,” Int. J. Oncol. 32(3), 653–662 (2008). [PubMed]
7. P. O. Andrade, R. A. Bitar, K. Yassoyama, H. Martinho, A. M. Santo, P. M. Bruno, and A. A. Martin, “Study of normal colorectal tissue by FT-Raman spectroscopy,” Anal. Bioanal. Chem. 387(5), 1643–1648 (2007). [CrossRef] [PubMed]
8. S. A. Boppart, “Optical coherence tomography - Principles applications and advances,” Minerva Biotecnol 16, 211–237 (2004).
9. A. F. Fercher, W. Drexler, C. K. Hitzenberger, and T. Lasser, “Optical coherence tomography - principles and applications,” Rep. Prog. Phys. 66(2), 239–303 (2003). [CrossRef]
10. M. Bhattacharjee, P. C. Ashok, K. D. Rao, S. K. Majumder, Y. Verma, and P. K. Gupta, “Binary tissue classification studies on resected human breast tissues using optical coherence tomography images,” J Innovat. Opt. Health Sci. 4(01), 59–66 (2011). [CrossRef]
11. K. W. Gossage, T. S. Tkaczyk, J. J. Rodriguez, and J. K. Barton, “Texture analysis of optical coherence tomography images: feasibility for tissue classification,” J. Biomed. Opt. 8(3), 570–575 (2003). [CrossRef] [PubMed]
12. C. A. Patil, N. Bosschaart, M. D. Keller, T. G. van Leeuwen, and A. Mahadevan-Jansen, “Combined Raman spectroscopy and optical coherence tomography device for tissue characterization,” Opt. Lett. 33(10), 1135–1137 (2008). [CrossRef] [PubMed]
13. J. W. Evans, R. J. Zawadzki, R. Liu, J. W. Chan, S. M. Lane, and J. S. Werner, “Optical coherence tomography and Raman spectroscopy of the ex-vivo retina,” J Biophotonics 2(6-7), 398–406 (2009). [CrossRef] [PubMed]
14. K. M. Khan, H. Krishna, S. K. Majumder, K. D. Rao, and P. K. Gupta, “Depth-sensitive Raman spectroscopy combined with optical coherence tomography for layered tissue analysis,” J. Biophot. (2013).
15. M. Larraona-Puy, A. Ghita, A. Zoladek, W. Perkins, S. Varma, I. H. Leach, A. A. Koloydenko, H. Williams, and I. Notingher, “Development of Raman microspectroscopy for automated detection and imaging of basal cell carcinoma,” J. Biomed. Opt. 14(5), 054031 (2009). [CrossRef] [PubMed]
16. N. Krstajić, C. T. A. Brown, K. Dholakia, and M. E. Giardini, “Tissue surface as the reference arm in Fourier domain optical coherence tomography,” J. Biomed. Opt. 17(7), 071305 (2012). [CrossRef] [PubMed]
17. A. F. Fercher, C. K. Hitzenberger, G. Kamp, and S. Y. El-Zaiat, “Measurement of intraocular distances by backscattering spectral interferometry,” Opt. Commun. 117(1-2), 43–48 (1995). [CrossRef]
19. G. Zonios, L. T. Perelman, V. M. Backman, R. Manoharan, M. Fitzmaurice, J. Van Dam, and M. S. Feld, “Diffuse reflectance spectroscopy of human adenomatous colon polyps in vivo,” Appl. Opt. 38(31), 6628–6637 (1999). [CrossRef] [PubMed]