Stimulated Raman scattering microscopy for rapid brain tumor histology

Yifan Yang*, Lingchao Chen and Minbiao Ji*‡ *State Key Laboratory of Surface Physics and Department of Physics Key Laboratory of Micro and Nano Photonic Structures (Ministry of Education) Collaborative Innovation Center of Genetics and Development Fudan University, Shanghai 200433, P. R. China Department of Neurosurgery, Huashan Hospital Fudan University, Shanghai 200040, P. R. China minbiaoj@fudan.edu.cn


Introduction
Although brain tumors only account for approximately 3% among all cancer types, they stand an important place in cancer research due to the fact that tumors of the central nervous system are second prevalent in children and adolescents only inferior to leukemia. 1 Among the brain tumors, gliomas carry one of the most dismal prognoses, with a two-year survival of about 2% for patients > 65 and 30% for those < 45 years.Up to date, surgery is still the most e±cient treatment for brain tumors.An optimal surgical management is associated with prolonging survival time and improving neurologic functions. 2owever, there remain several challenges in brain tumor resections.First, neoplastic tissues are often indistinguishable from healthy tissues during surgeries.Bright-¯eld neurosurgical microscopes are hard to visualize brain tumor margins and invasive cancer cells.Hence, tumor is often left behind and more than 85% of glioblastoma (GBM) recurrences occur at the resection margins, resulting in treatment failure and poor outcome. 3,4Conversely, unnecessary removal of normal brain tissues that do not contain cancer cells would generate neurological de¯cits. 5,6Second, a precise assignment of the tumor type and grade is signi¯cant because ¯rst-rank medical treatment varies widely depending on histological subtypes.Although some tumors of the central nervous system have their own individual forms of gross appearance, most of them are indistinguishable from each other. 7The importance of intraoperative histological classi¯cation during tumor surgery has been recognized for nearly 85 years. 3Third, diverse neuropathologies such as demyelination, primary CNS lymphoma or radiation injury caused by treatment of radiotherapy to known tumors are di±cult to be distinguished from glioma by magnetic resonance imaging (MRI) or computed tomography (CT). 8,9Thus, the risks of surgical resection are high for patients with aforementioned non-neoplastic conditions.
Hematoxylin and eosin (H&E) staining has been used as the gold standard histopathology method to diagnose brain tumors. 5Eosin stains proteins and cytoplasm bright pink, while hematoxylin stains basophilic structures (such as DNAs) blue-purple.However, H&E is a slow process that requires biopsy, ¯xation or freezing, thin sectioning and staining, thus cannot serve as an intraoperative tool.Even the intraoperative frozen sectioning, staining and histopathological consultation are still time-and labor-intensive ($ 30 min) and may delay clinical decision-making during surgery.Therefore, imaging methods delivering rapid and accurate diagnosis are highly desired for brain tumor resections.

Current Technologies
Advances have been made using a range of techniques to detect brain tumors to complement traditional histology.Noninvasive approaches for brain imaging to be stated here are CT, MRI, and positron emission tomography (PET), 10 while they are largely limited by lower spatial resolution and intraoperative compatibility.Preoperative MRI coregistered to patients has been used as a navigational guide to surgeries.However, it is unable to detect the full extent of cancer cell invasion and di®erentiate between radiation injury and recurrent malignancy. 11The shift of brain tissues during surgeries further reduces the spatial accuracy. 12Intraoperative MRI has shown great potential in providing updated images during surgeries, but is limited by its high cost and prolonged surgical duration. 13Fluoro-ethyl-tyrosine positron emission tomography (FET-PET) has been developed and tested as a tool to guide the resection of grade 3 to 4 gliomas with 88% sensitivity, but with only 54% sensitivity for lower-grade gliomas. 14Ultrasound (US) and optical coherence tomography (OCT) have been shown to provide structural information in real time.However, US is only accessible to large scale and di±cult to detect microscopic invasion, 15 and OCT lacks molecular speci¯city. 16 range of optical microscopy technologies has been developed and applied for intraoperative imaging.Fluorescence imaging of GBM involves labeling with °uorescing molecules such as °uorescein, protoporphyrin IX or 5-aminolevulinic acid (5-ALA) which is promising to reveal brain tumor margins, yet su®ers from several restrictions. 2,17,18irst, only $ 80% of the cancer cells absorbed the °uorescing molecules. 2Second, the nonspeci¯c labeling of dyes does not allow for tumor typing.Third, °uorophores may tend to photobleach under laser irradiations.Confocal microscopes have been applied for intraoperative imaging of tissues with °uorescence labeling, yet bare similar limitations of °uorescence imaging. 17Nonlinear optical imaging techniques have also been applied in imaging brain tissues.Second harmonic generation (SHG) microscopy is only speci¯c to noncentral symmetric contents such as collagen ¯bers and microtubules 19 and third harmonic generation (THG) microscopy is sensitive to refractive index inhomogeneity, but could not provide enough molecular information. 20he emerging platform for in vivo vibrational spectroscopic imaging delivers a new way to address these problems in current neuropathology and neurosurgery.2][23] Fourier transform infrared (FTIR) is able to perform chemical imaging of malignant gliomas and identifying brain tumor metastases. 24,25However, intraoperative IR microscopy is hampered by the strong water absorption and low spatial resolution given the longer wavelength ($ 2:5-10 m). 26Raman spectroscopy of brain tumors have been widely applied in rodent models and ex vivo human brain tissues. 27,28However, spontaneous Raman imaging of biological tissues has been limited by the weak signal intensities and slow imaging speed.Although handheld Raman probe demonstrated the capability in distinguishing tumor-in¯ltrated from normal brain, 29 it only provides spectroscopic information that relies heavily on spectral and statistical analysis, and introduces more complexity for clinical diagnosis.Surface enhanced Raman scattering (SERS)-based methods have the advantages of signi¯cantly enhanced Raman signals during in vivo imaging, but they require exogenous labeling with nanoparticles. 30

Basic Principles of SRS Microscopy
These limitations can be largely overcome by using coherent Raman scattering (CRS) microscopy, including coherent anti-Stokes Raman scattering (CARS) microscopy and stimulated Raman scattering (SRS) microscopy.In CRS, two excitation laser pulses, denoted as pump (!p) and Stokes (!s) interact with samples.Their di®erence (beat) frequency ¼ !p À !s matches the Raman-active molecular vibrations, thus coherently drives the corresponding chemical bonds and results in e±cient signal generation that allows up to video-rate rapid imaging. 31,32In CARS, the anti-Stokes photons are generated at ! as ¼ 2!p À !s via a wave-mixing process [Fig.1(a)]. 33Although CARS experiences resonance ampli¯cation with Raman modes, the signal persists even when !p À !s is detuned away from resonance.This so-called \nonresonant background" is originated from the electronic responses of materials and adds to the real part of the third-order optical susceptibility, which distorts CARS spectra from Raman spectra and has troubled researchers for around a decade. 34,35RS is analogous to stimulated emission process [Fig.1(a)] and was reported back in 1962. 36A few groups made early e®orts to integrate SRS spectroscopy with microscopy. 37,38In 2008, Freudiger et al. developed SRS microscopy with high frequency modulation and lock-in detection, and since then SRS microscope has become a rapidly growing tool for label-free chemical imaging in biomedical researches. 39SRS is known to have several advantages over CARS.First, the issue of nonresonant background was solved by SRS through self-heterodyne detection, probing only the imaginary part of the nonlinear susceptibility.SRS process results in the gain of Stokes photons (stimulated Raman gain, SRG) and a simultaneous loss of pump photons (stimulated Raman loss, SRL), and the net energy di®erence is converted to the vibrational energy in the excited states.Therefore, SRS could only occur at vibrational resonances [Fig.1(b)].Second, SRS spectra closely resemble those of spontaneous Raman scatterings without any distortion.Third, SRS signal is linearly proportional to chemical concentration, allowing easier quantitative analysis of complex system with di®erent chemical components. 40Hence, multicolor SRS could be applied to quantitatively analyze multichemical components (such as lipids and proteins), and provide virtual histopathology of biological tissues similar to H&E staining.[43]

Instrumentations of SRS Microscopy
The basic optical setups of SRS microscopes are shown in Fig. 2. 42,43 The tunable pump beam is directed from the output of an optical parametric oscillator (OPO) and the Stokes beam is the fundamental beam of the laser (e.g., 1064 nm).Both beams are overlapped spatially and temporally, and interact with the samples through a laser scanning microscope.In most cases, the Stokes beam is modulated at radio frequencies (1-20 MHz) where the noise spectrum of the laser has reached a shotnoise °oor.After the sample, the pump beam is ¯ltered out and sent to the photodiode (PD), and the SRL signal is demodulated by a lock-in ampli-¯er.SRS imaging can be performed in both transmission [Fig.2(a)] and re°ection (epi) modes.The transmission mode is good at imaging thin tissue sections and cultured cells, whereas the epi-mode has the advantage of imaging thick tissues and live animals. 32A specially designed annular detector was used to maximize photon collection e±ciency in the epi-mode, so that SRS imaging on live animals could be performed with video rate, 32 and in vivo brain tumor detection could be realized. 42ltiple choices of laser sources have been used for SRS microscopy.Most common lab-based SRS microscopes are equipped with narrowband picosecond laser sources, since they readily provide high spectral resolution (4-8 cm À1 Þ and sensitivity. 44,457][48] Furthermore, synchronized picosecond and femtosecond sources could be applied to realize hyperspectral SRS with parallel detection. 49,50Other types of multicolor and hyperspectral SRS could be found in previous works. 51wo-color SRS has been proven useful in brain tumor imaging with high sensitivity and speci¯city. 42,43Selective imaging at the CH stretching modes of 2845 cm À1 and 2930 cm À1 allows the di®erentiation of lipids and proteins [Fig.2(b)].Numerical decomposition algorithms could be applied to generate two-color maps of their distributions, forming histological images of biological tissues. 42,44However, current lab-based setups are still limited for clinical applications.Two recent technical advances have pushed SRS forward in the direction of clinical translation.First, ¯ber lasers with balanced detectors were developed for high quality SRS imaging, 52 thus bedside two-color SRS microscopes are available for rapid histology in the operating room [Fig.2(c)]. 41Second, simultaneous two-color SRS imaging method has been developed with dual-phase lock-in detection, allowing real-time virtual histology without the need of post processing. 57Such a dual-phase modulation and detection method has also been applied for other imaging modalities such as transient absorption microscopy. 53,54Further developments will include the miniaturization of the microscopes into handheld devices and endoscopes for in vivo applications.architectures and reveal clear di®erences between normal and tumorous regions of a GBM xenograft mouse brain.Furthermore, since SRS is a label-free method, it has the unique capability of in vivo imaging on live animals.An in¯ltrative human GBM xenograft mouse model was used to recapitulate the margin of human gliomas in the cortical surface and imaged through a cranial window with epi-SRS.The regions of brain tissues that appeared grossly normal under bright-¯eld microscope demonstrated extensive tumor in¯ltration with clear margins under SRS microscope based on both the chemical di®erences and histoarchitectures [Fig.3(c)].

Applications in Brain Tumor Detection
To assess the concordance between SRS and H&E images, web-based surveys were used to present large amount of SRS and H&E images in a random order to several certi¯ed neuropathologists, who were asked to classify these images into categories just as in the conventional practice. 41,43The survey demonstrates a remarkable concordance (Cohen's kappa) between SRS and H&E, proving SRS as an e®ective diagnostic tool for brain tumor.This method has been used in studies of both animal models and human surgical specimens.

SRS imaging of human brain surgical specimens
To evaluate the ability of SRS microscopy in detecting human brain tumors, surgical specimens were used to perform quantitative ex vivo studies. 43ormal histoarchitectures of human brain tissues were obtained from di®erent epilepsy patients, demonstrating SRS images of the gray and white matter junction [Fig.4(a)].As in xenograft models, while the white matters of human brain tissues contain large amount of lipid-rich myelin sheath, the gray matters possess less lipids and appear enhanced contrast of protein-rich cell nucleus.Frozen sections containing the gray-white junctions were also used to verify the origin of image contrast with combined SRS, H&E and Luxol fast blue (LFB) staining [Fig.4(b)].
The hallmarks of GBM in human brain tissues were studied with paired SRS and H&E.As in the mice model, these diagnostic information include hypercellularity, nuclear atypia of viable tumor, and mitotic features (Fig. 5). 43Lu et al. further reported that SRS could be used to discriminate necrosis from viable tumor [Fig.6(a)], evaluate vascular proliferation [Fig.6(b)], and reveal massive collagen deposition [Fig.6(c)]. 56SRS o®ers major advantages over other clinically available strategies conceived to detect tumor margins as it could simultaneously image proteins, lipids, blood and components of acellular regions by tracking only a few Raman frequencies.
Retrospective and prospective researches based on postoperative imaging have shown that in¯ltrating glioma cells adjacent to tumor core directly give rise to recurrence in approximately 90% of cases and a®ect progression-free survival.A complete resection based on intraoperative detection of in¯ltrating gliomas by MRI and US systems is currently impossible due to the limited spatial resolution.In contrast, SRS microscopy could reveal a gradual decrease in cellularity at the margins of both high-grade and low-grade gliomas [Fig.7(a)]. 43) brain tissues including gliosis and macrophage in-¯ltration.41 SRS is also well-suited for highlighting key di®erences in cellularity, microvascular proliferation and nuclear architecture that distinguish low-grade from high-grade gliomas [Fig.

Integration with machine learning for decision making
Although SRS microscopy might provide an accurate, powerful and rapidly obtained interpretation of histopathologic images, it still requires the expertise of pathologists, thus remains time-intensive and prone to inter-observer variability.Automated means that ensure rapid delivering robust and consistent diagnoses would greatly incorporate SRS microscopy into the existing brain-tumor surgery work°ow.A classifying program was developed based on the combination of cellular density, axonal density and protein/lipid ratio to rate the possibility of tumor occurrence within a FOV. 43Orringer et al. has recently employed a machine-learning algorithm for a more robust computational processing with multilayer perception neural network. 41The trained classi¯er could predict the probabilities that a given SRS image belongs to one of the four critical diagnostic classes: nonlesional, low-grade glial, highgrade glial or nonglial tumor.Machine-learningbased tissue diagnosis is expected to greatly bene¯t decision-making during operation, and will revolutionize intraoperative diagnosis combined with novel imaging methods.

Summary
In summary, we have reviewed recent progresses in developing SRS microscopy as a label-free digital histology tool for imaging brain tissues.Two-color SRS detection of lipids and proteins has demonstrated success in di®erentiating tumor from normal brain tissues, based on both the chemical and structural contrasts.A broad range of brain tissue types has been investigated, including normal brain, di®erent grades of gliomas, meningioma, oligodendroglioma and metastasis, with various diagnostic features of brain tumors.Machine-learning-based classi¯ers have been developed to provide more rapid and accurate diagnosis of these subtypes of brain tumors, pushing SRS further toward a practical tool for clinical translations.

Fig. 1 .
Fig. 1.Principles of CRS: (a) Energy diagrams of CARS and SRS, and (b) intensity and frequency changes in CARS and SRS processes.

Fig. 2 .
Fig. 2. Apparatus of SRS microscope: (a) Optical layouts of a typical SRS microscope, (b) Raman spectra of brain tissues in the regions of white matter, gray matter and tumor and (c) ¯ber laser-based SRS microscope that could be used in the operating room.EOM: electro-optical modulator, DM: dichroic mirror, PD: photodiode, GM: galvo mirrors, FL: optical ¯lter.

43 5. 1 .
Fig. 3. Two-color SRS images of human GBM xenograft mouse brains: (a) SRS and H&E images of a normal brain frozen section, (b) SRS and H&E images of a GBM in¯ltrated brain frozen section and (c) in the ¯eld of view where bright ¯eld microscope appears grossly normal, SRS image shows a distinct margin between tumor and normal brains.

Fig. 4 .Fig. 5 .Fig. 7 .Fig. 6 .
Fig. 4. SRS images of normal human brain surgical tissues: (a) Fresh tissue with a transition from white to gray matter and (b) gray-white junction imaged with frozen tissue sections, compared with H&E and LFB staining.