Differential Diagnosis of Lung Carcinoma With Coherent Anti-Stokes Raman Scattering Imaging

Coherent anti-Stokes Raman scattering (CARS) microscopy was first reported in the 1980s by a group at the Navy Research Laboratory.^29,30  It captures intrinsic molecular vibrations to create optical contrast.^31,32  In the CARS process, a pump field (ω_p), a Stokes field (ω_s), and a probe field (ω_p′, typically the same as ω_p) interact with the samples through a four-wave mixing process.³³  When the frequency difference, ω_p − ω_s (beating frequency), is in resonance with a molecular vibration, an enhanced signal at the anti-Stokes frequency, ω_as = ω_p − ω_p + ω_p′, is generated in a direction determined by the phase-matching conditions.³¹  The major advantage of CARS is that the signal yield is much higher, typically several orders of magnitude, than the signal yield obtained through the spontaneous Raman scattering process.³⁴  The intensity of CARS depends nonlinearly on 2 incident intensities: I_CARS ∼ I_p²·I_S.³⁵  As such, high peak powers are necessary and CARS signals are tightly restricted to the focal plane, allowing a 3-dimensional (3-D) segmentation capability,³⁶  similar to 2-photon microscopy.³⁷  In CARS excitation, the pump and Stokes fields coherently drive all resonant oscillators in the excitation volume at ω_p − ω_s, generating a third-order polarization at the anti-Stokes frequency.³⁸ 

The capability of generating optical contrast from endogenous chemical structures has led to a rapid expansion of CARS microscopy for high-resolution imaging in biomedical research. It has been used to visualize various tissue structures, such as skin,³⁶  DNA,³⁸  and retina.³⁹  In addition, CARS has been used in the imaging of brain sections.³³  Using the CH₂ stretching frequency, brain anatomy was revealed quite well at micron-level resolution.

Collectively, CARS microscopy offers many advantages. First, it generates optical contrast using molecular vibrations, which are intrinsic features of the specimen. Because no natural or artificial fluorescent probes are required, CARS effectively avoids the toxicity, photobleaching, and artifacts associated with the staining process. By tuning the beating frequency, CARS provides chemically selective excitation of characteristic vibrational resonances, allowing imaging of particular chemical structures at will. Second, because of its coherent nature, CARS microscopy is several orders of magnitude more sensitive than spontaneous Raman microscopy. As a result, CARS requires a much lower average power for excitation (∼0.1 mW), minimizing the phototoxicity caused by the imaging process and enabling its long-term in vivo imaging applications. Third, as nonlinear microscopy, CARS offers 3-D sectioning capability because CARS excitation is restricted to the focal volume. Combined with a fast scanning platform, this feature allows imaging of tissue and cell structures on a 3-D scale in real time. Finally, the optimum light sources for CARS are picosecond-pulse lasers. Compared with the use of femtosecond lasers, this feature reduces the chance of 2-photon–induced photo-damage and increases the light penetration depth in thick tissue.

Currently, in order to accurately delineate the type of lesions for definitive treatment, pathologists routinely stain lung biopsy tissue to examine changes in cellular and histologic features such as cell size, cell-cell distance, and formation of fibrous structures.⁴⁰  However, although this method is subject to interobserver variations, the CARS technique already provides high-resolution images that can clearly detect these features, without tissue staining with exogenous agents. Therefore, a pattern recognition method was developed in conjugation with CARS whereby such images were used as a basis for the quantitative classification of these cellular features in a way that would lead to a differential analysis of lung cancer.⁴¹  A total of 1014 CARS images were acquired from 92 fresh-frozen lung tissue samples. The established pathologic workup and cellular diagnostics were used as prior knowledge for establishment of a knowledge-based CARS system using a machine learning approach. This system functioned to separate normal, nonneoplastic, and subtypes of lung cancer tissues based on extracted quantitative features describing fibrils and cell morphology. Figure 1 shows representative CARS images and corresponding hematoxylin and eosin stain results of normal tissue, cancer, and noncancerous lesions. The normal lung is predominantly composed of well-organized fibrous structures, consisting of the bronchi and supporting matrix for alveoli (Figure 1, A and B). Cancer regions showed much denser cellularity compared with normal regions, and the size and configuration of the cells corresponded with these parameters, as shown by hematoxylin and eosin stain (Figure 1, C through H). Meanwhile, organizing pneumonia and interstitial fibrosis (Figure 1, I through L), 2 types of noncancerous lesions serving as controls, showed dense fibrous structures with distortion of normal lung architecture, similar to those shown by the corresponding hematoxylin and eosin stains.

Fibril and cell structures were used to separate cancerous from benign and normal samples because changes in these structures are closely related to different kinds of lung lesions, including cancer, pneumonia, and interstitial fibrosis.⁴⁰  In contrast, cancerous samples show a high density of cancer cells whose nuclei can be identified by CARS because of their low CH₂ level (Figure 1). Therefore, a computerized segmentation algorithm was used to precisely delineate the boundaries of cell nuclei.⁴¹  Measurements of morphologic characteristics, including nuclear size, cell volume, and cell-cell distance, were performed after segmentation for classification analysis of cancer subtypes. Figure 2 illustrates the 3D spatial distribution of normal, benign, and cancer samples (Figure 2, A), as well as cancer subtypes (Figure 2, B) using partial least square regression analysis, in which most of the disease groups are well separated. A support vector machine with recursive feature elimination–based classification analysis further showed the system's ability to distinguish lung cancer from normal and nonneoplastic lung tissue with 91% sensitivity and 92% specificity. Small cell carcinomas were distinguished from non–small cell carcinomas with 100% sensitivity and specificity. As an adjunct to submitting tissue samples to routine pathology, this novel system recognizes the patterns of fibril and cell morphology, enabling medical practitioners to perform differential diagnosis of lung lesions in mere minutes.

In spite of this progress, the developed platform showed only limited accuracy (around 70%–75%⁴²) for differentiating between 2 major cell types of non–small cell lung carcinoma, adenocarcinoma and squamous cell carcinoma. This difficulty is not surprising, and it corresponds with the clinical difficulty in differentiating these 2 cell types using morphology alone.⁴³  However, the advent of molecularly targeted therapies makes identification of the various histologic cell types and subtypes of lung cancer more important. Definitive diagnosis of these cell types is increasingly required for effective molecular treatment.⁴⁴  As a result, an efficient diagnostic strategy using small biopsies or cytology specimens, with minimum tissue consumption, will greatly benefit additional molecular tests that are frequently involved in reaching a definitive diagnosis for targeted therapies.^45,46  Accordingly, the potential of CARS has been further exploited by extending the previously developed 2-dimensional (2-D) image analysis framework into a 3-D nuclear segmentation, feature extraction, and classification platform. Using appropriate reconstruction and interpolation methods, a z-stack (a series of images acquired from different imaging depths) was modeled into a 3-D volume. Similar pathology-related features were measured in 3-D instead of 2-D to reach improved accuracy by eliminating certain 2-D artifacts. Thus, a 3-D measurement would provide access to the nucleus as a whole and allow measuring real nuclear size represented by volume.

In the work by Gao et al,⁴⁷  different mouse lung cancer models were developed to build a 3-D platform for nuclear segmentation and cancer cell type classification. Superpixel-based local clustering was used to perform segmentation and produce clusters that capture actual image boundaries with great accuracy in 2-D images⁴⁸  and was then extended to 3-D images, in which the clusters are called supervoxels. A modification of the simple linear iterative clustering algorithm was designed to compute 3-D supervoxels.⁴⁹  In this clustering process, the 3-D volume is partitioned into small regions based on local affinity information and voxels are grouped into one cluster according to the similarity of their features. The clustered data set was used to extract pathologically relevant features about the cell nuclei, which were then used to perform classification of lung cancer cell types. Figure 3 illustrates a representative case of the fitting results on all nuclei in an adenocarcinoma volume, where all ellipsoids are plotted with their color representing different depths within the z-stack. In Figure 4, comparative results of 2-D and 3-D classification methods are presented, where each point represents one calculating sample, specifically the one that was left out during a leave-one-out iteration. Classification from 2-D data (Figure 4, A) resulted in true-positive rates of 71.98% and 65.05% for adenocarcinoma and squamous cell carcinoma, respectively. False-positive rates, in the same order, were 28.02% and 34.95%. Classification results from the 3-D data analysis are plotted in Figure 4, B, where clear separation between cell types allows for data visualization on the same graph. Quantified classification accuracies were a 99.56% and 97.78% true-positive rate for adenocarcinoma and squamous cell carcinoma, respectively. False-positive classifications were 0.45% and 2.22%. The results showed that features extracted from 3-D data analysis provided information that significantly enhanced the classification accuracy and thus demonstrated proof of concept that information extracted from 3-D image analysis and segmentation can be used for the automatic diagnosis of lung cancer subtypes.

Like conventional optical imaging technologies, CARS faces the barrier of limited imaging depth caused by strong scattering inside the specimen; thus, the imaging depth is typically limited to less than 1 mm. Therefore, the development of CARS endoscopy is required for deep tissue imaging. A flexible endoscopy system consists of a light source unit, a detection unit, a control and display unit, and a fiber probe;^50,51  the latter is commercially unavailable for constructing a CARS endoscope. Compared with CARS, significant progress has already been made in developing 2-photon excitation fluorescence fiber microendoscopy.^52–57  The lag existing between the two is because 2 excitation lasers are required for CARS excitation whereas only 1 is needed for 2-photon excitation fluorescence. The additional laser beam creates many challenges, including efficient delivery of 2 ultrafast lasers using optical fibers, efficient collection of CARS signals, minimization of nonlinear optical effects, efficient design of chromatic aberration–corrected miniature optics, and miniaturization of laser scanning mechanisms.^58–61 

Several groups have reported progress toward developing optical-fiber–based CARS endoscopy systems. The first benchtop CARS fiber probe prototype was demonstrated by researchers at Harvard University in 2006.⁵⁸  They illustrated a proof-of-principle demonstration of CARS endoscopy by using a single-mode fiber (SMF) with a focusing lens attached to the fiber tip. They used a 2-dimensional translational stage to raster scan the sample and achieved imaging of polystyrene beads. To address the strong nonlinear effect inside fibers, that is, the four-wave–mixing (FWM) effect, researchers⁶²  at the University of California, Irvine, proposed a fiber probe design that incorporates 2 separate fibers for excitation light delivery and for signal detection in 2010. They investigated the probe design challenges, including capturing the backscattered CARS signals in the sample and the nonlinear effects of fibers for delivering the laser pulses. They illustrated that different biological tissues could be imaged ex vivo by using their fiber-delivered probe for CARS imaging. In addition, a new CARS probe design and its optical modeling results were reported by researchers at the University of Washington, INO, and University Health Network in Canada in 2010.⁶¹  They illustrated a miniature microscope objective design with a new spiral scanning fiber imaging technology and evaluated its anticipated performance through optical modeling. Their technique has the potential to make a CARS probe with a small footprint, an adjustable field of view, and high spatial resolution. The closest design to a real CARS endoscope to date is a laboratory prototype built by researchers⁶³  at Harvard University (Cambridge, Massachusetts) and the University of Washington (Seattle) in 2011. The fiber tip of the endoscope was driven at its mechanical resonance frequency with a spiral scan pattern by a piezoelectric actuator. The photodetector with optical filters was innovatively attached at the fiber tip to circumvent some serious issues induced by optical fibers, for example, background noises and the fiber coupling issue. The laser beams from the fiber output were focused onto the sample by a composite gradient index objective lens. This design has successfully demonstrated the capability of using the endoscopy prototype for in situ biological tissue imaging.

To date, we have made progress addressing these CARS endoscopy issues.^64–66  First, we demonstrated that a single multimode fiber can be used for both laser delivery and signal collection with better efficiency and reduced nonlinear effects in fibers than commonly used SMF. Although SMF has been successfully demonstrated for use in the CARS imaging system, its relatively small core makes SMF susceptible to small coupling efficiencies and generating nonlinear effects, which will reshape the spectra of laser pulses. Compared with SMFs, multimode fibers have larger core diameters, larger numerical aperture, and larger coupling efficiency. As shown in Figure 5, we demonstrated that multimode fiber can be used for delivery of picosecond excitation lasers without any degradation of CARS image quality.

We implemented a polarization-based FWM suppression strategy that dramatically reduces the FWM effect in fibers.⁶⁵  The probe fibers serve to deliver the excitation lasers to samples and collect CARS signals back from samples. When pump and Stokes lasers travel simultaneously in the delivery fiber, strong FWM noise at the CARS frequency is generated, which overwhelms the CARS signals and severely interferes with image interpretation. Until recently, there was no effective solution for solving the FWM problem. However, we were able to develop a strategy that uses only 1 multimode polarization maintaining fiber for both delivery and collection, resulting in reduction of FWM noise by ∼99%. Meanwhile, a dual-wavelength wave plate was specially designed to gain the maximum CARS excitation efficiency at samples. Figure 6 shows CARS images of mouse skin (Figure 6, A) and liver tissues (Figure 6, B) using the polarization and dual-wavelength wave plate strategy in the fiber-delivered CARS to suppress FWM; the cellular structure of the tissues can be clearly observed. Furthermore, we constructed a benchtop CARS endoscopy prototype using the multimode fibers and polarization strategy to capture both good-quality CARS images and spectra as shown in Figure 7. Now we are working on refining the prototype into a miniaturized endoscopy system.

Rapid advances in optical imaging technologies have led to substantial contributions to advancing diagnostic strategies in cancer research. As a label-free technique, CARS has already shown substantial potential for intraoperative applications by delineating pathologically related cell morphology information. To fully use its diagnostic value, an effective imaging and image analysis platform has been developed and tested on several types of cancers including lung cancer.^{41,42,67–69}  The focus is to realize accurate evaluation of CARS imaging data in a repeatable manner, promoting the translation of this technology for clinical use. These reported strategies were based on extraction and calibration of a series of disease-related morphology features for diagnosis using a machine learning approach.

With continuing progress to development of CARS-based microendoscopy,^{58,59,61,63–66}  another benefit of the reported image quantitation system lies in its potential for in vivo differentiation of lung carcinoma cell types and possibly subtypes during biopsy. With the advent of molecularly targeted therapies, avoiding sampling errors to allow complete identification of the various histologic cell types and subtypes within a given lung cancer becomes even more important. Coherent anti-Stokes Raman scattering allows selecting biopsy sites within a mass or nodule to minimize sampling error. Blind aspiration biopsy or fluoroscopic guided biopsy may sample surrounding tissues rather than the cancer and may not sample different cell types or subtypes within a cancer; thus, they may not be representative of the cancer as a whole. In contrast, CARS will not only ensure sampling of all cell types and subtypes within a heterogeneous lung cancer for molecular analysis at the time of biopsy, but also facilitate the whole diagnostic process while providing more abundant biopsy samples of the correct tissues for biomarker studies.

The funding of this research was initiated and supported by the Department of Systems Medicine & Bioengineering of The Methodist Hospital Research Institute, Weill Cornell Medical College, and the TT & WF Chao Foundation and John S. Dunn Research Foundation to Dr Wong. The authors would like to thank members of translational biophotonics and image-guided therapy groups of the Department of Systems Medicine and Bioengineering at The Methodist Hospital Research Institute for their helpful discussions in this review, as well as Rebecca L. Danforth, PhD, for proofreading the article.