1. Introduction

Currently, 63-98% of patients with lung, breast, bladder, and colorectal cancer undergo surgery. For tumor resection, the status of surgical margins (involved or not by malignant cells) is considered to be one of the strongest indicators of success and patient outcomes [1]. Unfortunately, with 15–60% of these initial surgeries concluding with positive margins, patients will quite often have less favorable outcomes for their cancer, and often will have to undergo follow-up surgeries or therapies resulting in a negative impact on their overall experience and morbidity along with significantly increased health care costs [2]. Improving margin status assessment and detection in general, and especially during the surgical procedure would undoubtedly improve patient care and prognosis, while saving billions of dollars each year by drastically decreasing the need for additional surgeries or therapies. Further significant cost reductions may come from reducing the operating time required for margin assessment [3,4]. Assessing tumor margins is not always easy, as tumor tissue may not always be distinct from surrounding healthy tissue to the naked eye, a well-known issue in skin, breast and pancreatic cancers [5–7]. Current techniques to determine margin status of surgically resected tumors include visual inspection, and histopathological (microscopic) analysis of either frozen or formalin-fixed paraffin-embedded representative tissue sections, usually stained with hematoxylin and eosin (H&E). Regrettably, all these methods have severe limitations [8]. Visual inspection alone is often not adequate to demarcate boundaries between tumorous and healthy tissue and will miss microscopic disease extending to the margin. Histopathological frozen section analysis (FSA) can be used while the patient is in the operating suite but depends on visual inspection, will increase surgical times (∼20 minutes per section) and suffers from reduced accuracy depending on the clinical–pathological setting, with significant (from 20 to 36%) false negative rates in the determination of margin status in some instances, such as breast conserving surgery [9,10]. Pathological analysis of H&E-stained so-called ‘permanent’ sections obtained from formalin-fixed paraffin-embedded tissue is the most accurate of these procedures and is considered the clinical standard for routine management of patients. The significant downside of this procedure is that it has the longest turnaround time, requiring in general about 24 hours to get the samples fixed, embedded, cut, stained, and analyzed. Obviously, it is not suitable for feedback in real time during surgery. As such, typical wait times for histopathological analysis can be upwards of a week in cases where the specimen cannot be processed in-house. Limited sampling area from surgical specimens, and serial processing for each sample add additional dilemmas with this technique that contribute to the lengthy processing times and in some cases may decrease margin determination accuracy, given the partial sampling of the margin area [11,12]. Still, brightfield microscopic examination of selected, well-chosen, presumably representative H&E stained paraffin-embedded sections is considered the gold standard for margin determination by histopathological analysis and is the typical comparison that new, emerging technologies are evaluated against. These challenges indicate the need for a new detection technique to rapidly image fresh tissue samples during surgery to provide precise, rapid feedback for the surgeon to resect the entire tumor with negative margins.

Non-linear microscopy (NLM) has been used as a virtual histological imaging technique, enabling high resolution and high sensitivity images of tissue [10,13,14]. NLM techniques have been used to obtain contrast from fluorescent nuclear-bound stains such-as from acridine orange and sulforhodamine 101 via two-photon excitation along with a simultaneous label-free collagen imaging using second harmonic generation to demonstrate virtual H&E like images [10,15]. However, staining techniques typically have a range of staining variability as well as contamination of the sample itself. This, combined with varying interpretations that depend on individual histopathologists can lead to large differences in diagnosis interpretation on the same sample [16–18]. The addition of foreign contaminants to the sample can lead to the inability to conduct additional diagnostic tests or repeat imaging if closer inspection is required. A need for a label-free technique is of great importance to preserve the sample in its original form and to not limit further diagnostics. Furthermore, NLM has more complicated systems leading to a much higher upfront cost to purchase the necessary equipment such as ultrafast lasers, and high-speed multi-channel digitizers [10,13].

A label-free approach called optical resolution photoacoustic microscopy (OR-PAM) has recently demonstrated the ability to create label-free virtual H&E images but requires physical contact of an ultrasound transducer to the sample to measure the produced acoustic waves. This absorption contrast system focuses on measuring the photoacoustic effect with an ultrasound transducer. Cytochrome contrast has been realized with 420-nm excitation light and nuclei contrast with 266-nm [19,20] as well as with mid infra-red light to differentiate additional protein and lipid structure [21]. High resolution cellular images have demonstrated close concordance with conventional H&E staining as well as statistically significant diagnostics results with the current gold standard of H&E staining [19]. Unfortunately, the requirement of physical contact with the sample results in the inability to image in vivo samples as well as putting an upper limit on the use of high numerical apertures in reflection mode operation.

Photoacoustic remote sensing (PARS) is an emerging technology well suited for this application by demonstrating improvements over the aforementioned microscopy methods with a label-free approach that does not require contact with the sample. PARS is an all-optical reflection mode imaging system that uses the absorption of light to produce localized ultrasound waves that are optically detected with an interrogation beam. PARS has used absorption contrast to image biological components such as vasculature, SpO₂ levels, cell nuclei, lipids and cytokines using specific excitation wavelengths of light that are known to be readily absorbed by the molecules of interest [22–25]. Nanosecond excitation laser pulses heat the molecule of interest to produce a large pressure gradient which modulates the refractive index, typically detected by near infra-red (NIR), continuous wave non-coherent lasers. Recent work with ultraviolet-PARS (UV-PARS), has demonstrated excellent comparisons to hematoxylin stains by resolving nucleic acid contrast in cell nuclei [25–29]. Several approaches have begun to observe complimentary cytochrome staining using different wavelengths but suffer from limited low-repetition-rate laser sources and thus extended imaging times [23].

Our work demonstrated here uses a single excitation laser at 266-nm for the absorption contrast of cell nuclei as a virtual hematoxylin stain while simultaneously gathering the PARS interrogation beam scattering signal for a virtual eosin stain. We demonstrate the capabilities of this dual-contrast system for en-face 2D and 3D imaging of various tissue samples with high concordance with H&E staining procedures. Additionally, we use co-integrated confocal fluorescence microscopy (CFM) to validate cell nuclei imaging in 2D and 3D, while also demonstrating susceptibility of fluorescence methods to staining variability. To our knowledge, this is the first microscopy approach capable of depth-resolved imaging of unstained thick tissues with virtual H&E contrast. The system has potential for real-time virtual histological imaging of fresh tissues enabling future point-of-care applications with substantial improvements in patient care and prognostics.

2. Methods

2.1 Experimental setup

Figure 1 depicts the system diagram of UV-PARS/scattering microscopy with laser scanning CFM for the purpose of obtaining virtual H&E stains and validation with CFM. A 266-nm nanosecond excitation laser is pulsed on the sample to introduce rapid heating leading to a thermal expansion and a rapidly increasing pressure gradient, better known as the photoacoustic effect. This large pressure gradient will create modulation in the local refractive index of the sample which can be detected using a circularly polarized 1310-nm continuous-wave superluminescent diode source (Thorlabs, SLD1018PXL). The interrogation beam is combined using a harmonic beam splitter (Thorlabs, HBSY134), with the 266-nm laser, created through second harmonic generation by focusing the output of a 532-nm pulsed fiber laser (IPG Photonics, GLP-10) into a Caesium Lithium Borate (CLBO) (Eksma Optics) crystal. These two co-aligned beams are rapidly scanned with galvanometers (Thorlabs, GVS412) through a 0.5 numerical aperture (NA) reflective objective (Thorlabs, LMM-40X-UVV) onto the sample. Both the excitation and interrogation beams are shaped using Galilean beam expansion to fill the aperture of the reflective objective. The back-reflected interrogation beam is separated from the excitation laser with the harmonic beam splitter previously mentioned, then directed away from the source through the use of a polarizing beam splitter (PBS) (Thorlabs, CCM1-PBS254) to redirect the backscatter signal towards a 75-MHz balanced photodiode (Thorlabs, PDB420C-AC).

 

Fig. 1. System diagram of PARS and scattering microscopy cointegrated with laser scanning confocal fluorescence microscopy. Components can be identified as: avalanche photodiode (APD), balanced photodiode (BPD), beam dump (BD), beamsplitter (BS), Caesium lithium borate crystal for SHG (CLBO), collimator (C), dichroic mirror (DC), galvanometer mirrors (GM), half-wave plate (HWP), harmonic beam splitter (HB), mirror (M), lens (L), polarized beam splitter (PBS), prism (P), quarter-wave plate (QWP), and reflective objective (ROBJ).

Download Full Size | PDF

The confocal fluorescence subsystem used the 532-nm pulsed laser to excite the fluorophore, propidium iodide (PI), and an avalanche photodiode to detect the fluorescent emission, while operating in parallel with UV-PARS. PI has a maximal excitation peak at 535nm and maximal emission peak at 617nm. A portion of the 532-nm excitation signal is separated from the UV-PARS pathway using a 70:30 beamsplitter (Thorlabs, BST10) with the intensity of fluorescence excitation controlled using a continuous neutral density filter wheel (Thorlabs, NDC-50C-4M). The 532-nm excitation light is co-aligned into the UV-PARS system to be co-scanned by the galvanometers into the reflective objective. The fluorescence emission is redirected using a long pass 550nm dichroic filter (Thorlabs, DMLP550R) then further filtered using a 532-nm laser line filter (Thorlabs, FL533-17) and bandpass filter of 630nm/69nm (Thorlabs, MF630-69). The fluorescence emission is collimated (Thorlabs, F280FC-A) then fed into a single mode fiber (SMF) (Thorlabs, P5-460B-PCAPC-1) for spatial filtering to achieve confocality. The fiber is directly coupled to an avalanche photodiode for digital signal acquisition (Thorlabs, APD120A2).

Digital data was acquired from five channels using a 12-bit, 125MS/s data acquisition card (GaGe, CSE8389-2GS) with 32 samples recorded post trigger. This includes data streams for the UV-PARS signal from the RF output of the balanced photodiode, the fast axis and slow axis position of galvanometer scanning system, the fluorescence emission signal from the avalanche photodiode, and the DC scattering signal from the photodiode. Signal filtering for the RF signal included being passively bandpassed between 1.8MHz-22MHz while the DC scattering signal was low-passed at 100kHz. Data was processed by taking the max of the UV-PARS signal and the mean of the galvanometer, fluorescence emission and DC scatter signals. The sample was moved using an axial stage (Zaber, X-VSR-E) and X-Y motorized stage (Thorlabs, MLS203-1). Optimization of laser pulse spacing was computed by varying the slow and fast axis frequencies of the galvanometer. Unlike other implementations of PARS imaging, the slow axis moves along a ramp function rather than a sinusoid. This allows for a full-sweep of the FOV in a single period regardless of when the acquisition starts without duplication of any points on the image, thereby reducing optical exposure times and effectively doubling our imaging speed. The slow axis frequency is calculated as the ratio of the laser pulse repetition rate (PRR) and number of points (N) desired. The fast axis frequency was calculated as ${f_{fast}} = {f_{slow}}\sqrt {N/2} = \sqrt {PR{R^2}/({2N} )} $. Using a 40kHz PRR and 200,000 points per images as used in this study, this equates to slow and fast axis frequencies of 0.20Hz and 63.2Hz respectively and corresponds to an image acquisition time of 5s per single FOV.

2.2 Sample preparation

All animal samples were acquired in accordance with the University of Alberta’s Animal Care and Use Committee ethics guidelines and regulations. Human tissues were acquired using ethics protocol HREBA.CC-20-0145. Animal organ tissue sections were dissected from a nude mouse (Charles River, NU/NU) following euthanasia. Human tissue was obtained from surgical breast cancer resections. Tissues were washed with phosphate buffered saline and immersed in 10% neutral buffered formalin for fixation, embedded in paraffin blocks, and lastly sectioned to the desired thickness of 4µm or 30µm and adhered to a glass slide. Prior to imaging, deparaffinization was performed by first heating the slides at 60°C for 1 hour, followed by 2-minute-long washes in 2 changes of xylene, 2 changes of 100% ethanol, 95% ethanol, and finally deionized water. For fluorescence imaging, samples were washed 3 times with phosphate buffered saline and incubated with RNase (ThermoFisher, EN0531) for 60 minutes at room temperature to remove RNA from the cell. Next, samples were equilibrated in phosphate buffered saline, with 300µL of 500nM PI (Thermofisher: P1304MP) being added to the sample then incubated at room temperature for 5 minutes. Lastly, the sample was rinsed 3 times with phosphate buffered saline before imaging.

2.3 Li’s ICA

To quantify the concordance between virtual and true hematoxylin staining, we used Li’s intensity correlation analysis (ICA), a metric thought to be better suited to our application than Pearson’s correlation coefficient as it is robust to intensity scaling and across-image intensity variability. Li’s ICA value has a maximum of 0.5 for highly correlated images and −0.5 for high discordance [30].

3. Results

We demonstrate our co-integrated system’s ability to realize fluorescence emission and absorption contrast from cell nuclei, with complimentary scattering contrast from surrounding tissue in both thick and thin tissue samples as a means for virtual H&E staining. We compared virtual histology images of tissue sections with CFM and adjacent H&E-stained sections using our new microscopy system. Direct comparisons between CFM and UV-PARS are demonstrated in the same planar tissue section instead of using adjacent samples as is typical with H&E. Depth-resolved imaging of thick tissue was taken using UV-PARS and scattering contrast at depths up to 24µm.

3.1 Scattering lateral and axial resolution

Scattering contrast axial and lateral resolution was demonstrated for the 1310-nm non-coherent continuous wave laser using a USAF1951 airforce target and 7.2µm carbon fibers (AGM94CF0500). The airforce target was lowered in intervals of 3µm while images were acquired to create a depth-resolved data set. The intensity at each depth was graphed to create an edge spread function (ESF). The derivative of the ESF was taken to acquire a line spread function (LSF) whose FWHM was measured to determine the depth of focus (DOF) to be 26.5µm as shown in Fig. 2(a). Lateral resolution was realized by taking the FWHM of the ESF derivative at the edge of a carbon fiber. The lateral resolution of the scattering microscopy sub-system was found to be 2.4µm as seen in Fig. 2(b).

 

Fig. 2. Resolution characterization of 1310-nm scattering contrast beam. a) Axial resolution was determined to be 26.5µm by imaging the airforce target while translating the axial stage at 3µm intervals. b) Lateral resolution was determined to be 2.4µm using an edge spread function and its derivative with carbon fibers.

Download Full Size | PDF

The UV-PARS sub-system lateral and axial resolutions were previously characterized in Haven et al. to be 0.39µm and 1.2µm respectively [25]. When extrapolating the resolution of 0.39µm for 266-nm, it is expected for the 1310-nm wavelength to be 4.9 times greater when considering the Rayleigh Criteria. Our results demonstrate a slightly larger than expected increase of 6.15 times. This is expected due to the differing beam diameters of the two different wavelengths into the aperture of the reflective objective.

3.2 Tissue imaging

We show that our virtual H&E images provide close concordance with true H&E histology, by comparing virtual H&E images of thin unstained 4µm tissue sections with adjacent tissue sections processed with H&E labeling. We obtained 0.5mm x 0.5mm mosaic images of PI fluorescence labeled nuclei, UV-PARS, and combined virtual eosin-like staining from a 4µm thick section of a mouse lung with a clearly defined bronchiole surrounded by alveoli as shown in Fig. 3. The adjacent H&E section, Fig. 3(d) shows confirmation that both images have captured the cell nuclei contrast and display nuclei by nuclei agreement in the region of interest (ROI) with UV-PARS and fluorescent images demonstrating Li’s ICA value of 0.399. Furthermore, the signal-to-noise (SNR) of UV-PARS and CFM was 44dB and 28dB, respectively. Additional virtual eosin staining obtained via scattering contrast is compared to traditional histopathological staining with exemplary results, providing an entirely virtual H&E stain in a fraction of the time traditional H&E staining takes.

 

Fig. 3. A 4µm section of mouse lung tissue, 0.5 mm x 0.5 mm FOV. Scale bars are all 150µm. a) UV-PARS absorption contrast of cell nuclei. b) Propidium iodide fluorescence. c) UV-PARS (blue) with scattering microscopy contrast (pink) for false staining of cell nuclei contrast and protein contrast. d) H&E stain of adjacent 4µm section.

Download Full Size | PDF

Images were also obtained from the gastrointestinal tract to demonstrate our system’s imaging ability in a variety of tissues to highlight imaging versatility. Prominent villi with excellent correlation to both fluorescence and H&E stains are shown in Fig. 4. Images are 0.16mm x 0.5mm. Previous literature points to staining variability as a potential source of error in pathological diagnostics. To demonstrate the propensity of fluorescence staining variability and the potential improvement afforded by our label-free approach, we note that some cell nuclei present in UV-PARS images are absent in fluorescence images. We quantified signal intensity and concordance between fluorescence CFM, Fig. 4(b), and UV-PARS, Fig. 4(a), images of cell nuclei, using the SNR and Li’s ICA value. SNR was significantly higher in UV-PARS with an SNR of 41dB in Fig. 4(a) while CFM was just 24dB in Fig. 4(b). Li’s ICA was found to be 0.314 for Fig. 4, a lower than expected value. This is attributed to the additional contrast shown in the UV-PARS image from around the nuclei leading to less correlation between pixel intensities in the two images. Additionally, the lack of signal for all nucleotides in CFM would have also caused a decrease in the expected ICA value. Adjacent sectioning appears to have impacted the leftmost villus where it has poor signal in the UV-PARS and PI images. However, due to both UV-PARS and fluorescence images not showing the same nuclei, we can infer that the adjacent section did not contain the same structure as the H&E stained one. The addition of the virtual eosin contrast with UV-PARS in Fig. 4(c) is analogous to traditional paraffin-embedded H&E, Fig. 4(d), with the ability to digitally resolve the nuclei with even greater contrast. However, some of the additional contrast depicted in the UV-PARS image is from the surrounding cytoplasm, not the cell nuclei. Further work needs to be done to characterize this contrast in the image but residual nucleotides in the tissue may contribute to the background signal in the UV-PARS image.

 

Fig. 4. A 4µm section of mouse bowel tissue, 0.16 mm x 0.5 mm FOV. Scale bars are all 50µm. a) UV-PARS absorption contrast of cell nuclei. b) Propidium iodide fluorescence. c) UV-PARS (blue) with scattering microscopy contrast (pink) for false staining of cell nuclei contrast and protein contrast. d) H&E stain of adjacent 4µm section.

Download Full Size | PDF

Subsequent work has progressed to imaging Human breast samples with our work highlighted in Fig. 5 in where we observe a section of adipose tissue. Here we examine the alignment of nuclei with the scattering microscopy’s visualization of sectioned breast tissue. We observe higher concentrations of nuclei in the areas with more cytoplasm and nuclei at set intervals along the outer perimeter of the fat vacuoles. This image was acquired by constant velocity motor-stage translation in the x-direction and single axis galvanometer scanning in the y-direction, with a fast acquisition time of 10s over a FOV of 880µm x 180µm.

 

Fig. 5. Human breast tissue of adipose tissue. All scale bars are 100 µm. a) UV-PARS with nuclei absorption contrast. b) Virtual hematoxylin (blue) using UV-PARS and eosin (pink) using scattering microscopy.

Download Full Size | PDF

To demonstrate 3D virtual histology, we scanned thick unstained murine lung tissue using virtual H&E UV-PARS and scattering microscopy. We showcase 4 images at 6µm separated depths in Fig. 6(a). The maximum resolved depth was defined to be the maximum depth which we were able to resolve reasonable PARS signals, shown to be 24µm in Fig. 6.

 

Fig. 6. Depth-resolved 3D virtual H&E imaging of a thick tissue section of a mouse lung. Images were optically-sectioned at 6µm intervals. a) H&E virtual stain with UV-PARS and scattering microscopy. b) UV-PARS nuclei contrast in thick tissue. Scale bar is 50µm. c) CFM nuclei contrast in thick tissue. Scale bar is 50µm.

Download Full Size | PDF

These images highlight our systems ability to penetrate deeper into tissue and resolve single cell nuclei with complimentary virtual eosin contrast at each interval. With our data acquisition set-up and high repetition rate laser this depth-sectioned image can be acquired in 20s. We compare our UV-PARS, Fig. 6(b), with CFM, Fig. 6(c), to visualize nuclei by nuclei colocalization in thick tissue.

4. Discussion

Virtual H&E staining is realized using UV-PARS and scattering contrast in a fraction of the time the current staining protocols take to obtain images. Our system can produce two separate images simultaneously resolving the cell nuclei replicating the hematoxylin stain using the back-reflected AC signal and complimentary virtual eosin stain using the back-reflected DC scattering signal of our interrogation beam. Other photoacoustic, all-optical combined systems have been demonstrated such as OCT and PARS [31] but none yet which have demonstrated simultaneous virtual H&E. Virtual H&E imaging is demonstrated by providing gross imaging information of villi and a bronchiole that are comparable to adjacent slices of paraffin-embedded H&E stained tissue. These images have maintained a high degree of colocalization with the adjacent H&E sections as with the fluorescence of cell nuclei using CFM. Discrepancies in UV-PARS with the CFM images can be attributed to both staining variability leading to color and intensity variation and differences in their respective DOF [16–18].

By acquiring both the virtual eosin and hematoxylin signals separately, it is possible to perform individually optimized signal filtering and image reconstruction for both scattering contrast and the UV-PARS imaging modalities. The AC and DC signals are acquired separately, with individual bandpass and low-pass filtering implemented before digital signal processing. Our new method of filtering has enabled a completely new subset of digital signal to be acquired and processed, visualizing virtual eosin contrast. Further work is imperative to quantify the separate frequency domains to uncover if additional molecular contrast could be included or to further optimize our methods described in this manuscript in thicker tissues where scattering will have greater interference and may not be a viable option.

While it is well known that UV light can cause DNA damage, our results show the ability to take repeated scans with minimal visible degradation [32]. We, along with others, have previously shown that H&E staining was not adversely affected by first doing a UV-PARS scan [25]. Also, our recent work has shown that UV surface fluence to obtain UV-PARS images are below ANSI limits when focusing 13.3µm below the surface while using 5nJ pulse energies of UV light [33]. Here, a surface fluence of 2.7mJ/cm² is demonstrated, just below the ANSI skin maximum permissible exposure (MPE) limit of 3mJ/cm². New data also suggests sub-nanojoule pulse-energies of 750pJ could be used for UV-PARS [26], further paving the way for safe point of care imaging in both ex vivo and potentially in vivo applications [34].

UV-PARS has demonstrated a profound ability to image an expanding collection of different tissues including brain, gastrointestinal, breast and its use in Mohs surgery [26–29]. Currently, our imaging acquisition takes 3 minutes to acquire a 1mm x 1 mm section using galvanometer scanning or 1 minute for stage and galvanometer scanning, 2x faster than the scan-speed reported in Orringer et al. [31]. Our approach will still be slower than open-top light-sheet microscopy and nonlinear microscopy methods which can achieve fast imaging rates of 19fps, however, we will avoid the staining and tissue clearing steps which take several minutes [10,15,35]. In the future, with a faster laser repetition rate imaging rates could be considerably faster than the stimulated Raman scattering approaches as well as offer comparable imaging speed to NLM methods, all without tissue staining. For example, with a laser repetition rate of 2MHz, our imaging technique could potentially acquire cm x cm images in 4 minutes with a lateral resolution of 0.5µm.

The resolution of the scattering microscopy sub-system is currently limited by our 1310-nm wavelength and could in theory be greatly reduced by changing the interrogation wavelength. Our UV-PARS sub-system achieved 390nm lateral resolution, comparable to the 360nm lateral resolution reported with stimulated Raman microscopy [31]. Our reported UV-PARS resolution is finer than open-top light-sheet microscopy and comparable with nonlinear microscopy while being completely label-free.

Our imaging depth of only tens of microns is limited by UV penetration. However, this may be sufficient for many potential clinical applications such as en-face imaging of breadloafed lumpectomy specimens or imaging beneath the surface of non-sectioned lumpectomy tissues, where for example, margins of a few cellular layers is considered adequate for cancers such as invasive carcinoma.

Applying our new system could lift this imaging modality into more robust clinical applications with the ability for real-time visual feedback for the observer during tissue examination. Our virtual H&E imaging system is currently speed-limited by a 600kHz repetition-rate laser. While future work could explore higher repetition rates, our current approach can operate at pixel-readout rates over 24 times faster than recently published dual-wavelength PARS approaches for H&E imaging [29]. Moreover, such dual-wavelength methods have not yet demonstrated 3D imaging capabilities. With our current 600kHz repetition rates, we could in principle image a 2mm rim of a 1” diameter bread-loafed lumpectomy area in 6 minutes with the addition of our fast mechanical scanning translation stage. Local spot-checking of suspicious areas of resected tissues could be an achievable objective that may have high clinical significance for identifying positive margins in oncologic surgeries.