Expand this Topic clickable element to expand a topic
Skip to content
Optica Publishing Group

Dual-modality hyperspectral microscopy for transmission and fluorescence imaging

Open Access Open Access

Abstract

Conventional color imaging provides information in the red, green, and blue (RGB) channels with limited resolution in the spectral domain. This could lead to poor sensitivity in imaging biological samples. In fluorescence imaging, the use of multiple fluorophores is often necessary to label proteins and DNAs for in vivo experiments. Crosstalk between fluorophores can be troublesome because it is challenging to differentiate each fluorophore when their emission spectra are overlapped. To help address these issues, we developed a dual-modality hyperspectral microscopy system that combines hyperspectral imaging and microscopy imaging to provide spatial and spectral information of the samples. The dual-modality feature allows us to study biological samples and fluorescent samples using the same system. We show that applications of the system enable: the identification of different tissue and cell structures; identification of each quantum dot fluorophore, calculation of their relative proportions at each location, and elucidation of their spatial distributions in a mixture sample containing multiple types of quantum dot fluorophores. The results give a brief showcase of how hyperspectral imaging can be useful for biomedical imaging and fluorescence imaging applications.

© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement

1. Introduction

In biomedical imaging, histopathology is an important step in the diagnosis and study of diseases. It involves a pathologist examining the tissues and cells of a biopsy specimen on a glass slide under a microscope. Based on the color differences and patterns of tissue structures, the pathologists will be able to identify cell nuclei, cytoplasm, and extracellular matrix (ECM) and make a diagnosis on the type of cancer, its grade, and sometimes tumor margins for surgical operations. However, with a conventional white light microscope, although the images obtained show a color contrast, they are captured using three spectral bands, i.e. red, blue, and green (RGB), with limited spectral information. Hyperspectral imaging (HSI) is an imaging technique that combines the principles of optical imaging with spectroscopy and provides additional spectral information in hundreds of spectral bands. It saves the image data in a three-dimensional data cube including the spatial dimensions (x, y) and the spectral dimension (λ) [1]. As such, each pixel in a hyperspectral image contains the complete transmission, fluorescence, and/or reflection spectrum of the sample, and conversely, each spectral band contains its corresponding single-band image. It has recently attracted great attention in cancer diagnosis, skin wounds, diabetic foot, etc [15]. Many have also combined hyperspectral imaging with microscopy imaging for added functionalities [69].

In fluorescence imaging, although spatial information can be acquired about the localization of fluorophores across a sample, spectral information is needed to identify the type of fluorophore present in the sample based on its unique spectral fingerprint. Conventional fluorescent microscopy uses excitation and emission filter sets designed for specific fluorophores, thus allowing for one fluorophore to be imaged at a time. However, imaging multiple fluorophores with different spectral characteristics remains a tedious process, requiring the use of separate excitation and emission filters that are optimized for each fluorophore, which has to be swapped out before imaging each fluorophore; as such the sample has to be imaged sequentially for each fluorophore used. Unconventional fluorophores pose an added challenge when they do not match well with the spectral bandwidth of commonly used emission filters. Furthermore, it is challenging to identify fluorophores with highly similar emission spectra and large spectral overlap, i.e. crosstalk. This may impose limitations on the experimental design when multiple fluorophores are intended to be used simultaneously or when imaging targets in the sample are located close to each other spatially. Hyperspectral imaging is a promising solution because the spectral band selection is very precise allowing the detection of multiple fluorophores at one go in a short period of time [10,11]. It has also been demonstrated for fluorescent microscopy imaging of nanoparticles such as quantum dots [12,13].

In this paper, we present a dual-modality hyperspectral microscopy system as well as image analysis workflows aiming to give a brief overview of how hyperspectral imaging can be used for biomedical and fluorescence imaging applications. The dual-modality feature enables the observation of both transmission and fluorescent properties of samples using the same system. In particular, we show that for biomedical imaging, hyperspectral imaging allows the identification of different tissue and cell structures on human breast histopathology slides; and for fluorescence imaging, hyperspectral imaging allows the identification of each type of quantum dot fluorophore, quantification of their relative proportions, and elucidation of their spatial location in a quantum dot mixture environment with minimal crosstalk. Our results provide a good showcase of the advantages of a dual-modality hyperspectral microscopy system compared to a conventional white light microscopy system for transmission and/or reflection measurements and a conventional fluorescent microscopy system for multiplexed fluorescent measurements.

2. Methodology

2.1 Hyperspectral imaging microscopy system

The schematic drawings of the custom-built hyperspectral imaging microscopy system are shown in Fig. 1(a) and 1(b) for transmission modality and fluorescence modality, respectively. The microscope is upright. The key feature is that the configuration allows switching between modalities by only changing the light source position without changing other components of the microscope body. In the transmission modality, the light source is a broadband white light source (HPLS343, Thorlabs) and is connected to a condenser (Nikon 0.9 NA D-CUD air condenser) illuminating the sample from below. In the fluorescence modality, the light source is an ultraviolet (UV) LED light source (M375L4, Thorlabs) illuminating the sample from above. The detection beam path is shared by both modalities which includes an objective (Nikon Plan Fluorite objectives 4x/0.13, 10x/0.3, and 20x/0.5), a 50:50 beam splitter (working range: 350 - 1100 nm), a variable bandwidth liquid crystal tunable filter (Kurios-VB1, Thorlabs), a tube lens, and a monochromatic CMOS camera (CS126MU, Thorlabs). The incoming light after passing through the collimation optics is then focused onto the sample. The transmission or fluorescence signals of the sample are collected by an objective lens and pass through the liquid crystal tunable filter, which isolates the light of a narrow band of wavelengths corresponding to the desired spectral band. It has three selectable operating bandwidths of FWHM = 55 nm (wide), 21 nm (medium) and 10 nm (narrow). It contains no moving parts with a switching speed of less than 50 ms (wide), 80 ms (medium), and 200 ms (narrow), thus the optical pathway through the system is not affected during operation and the captured images are not subject to any undesirable pixel shift effect. The optical density ranges from 0.5 to 4.5 with the minimum value at 625 nm. The spectral range is 420 - 730 nm with a minimum step size of 1 nm resulting in a maximum possible 311 spectral bands that can be captured and the bandwidth is 18 nm for each spectral band. Each single-band image is then captured by the CMOS camera, which stores 12-bit grayscale images of a high resolution of up to 4096 × 3000 pixels at a maximum frame rate of 21.7 fps, with a pixel size of 3.45 µm x 3.45 µm. Finally, all single-band images from individual spectral scans are stacked together to form a hyperspectral image. In all our experiments, all 311 spectral bands were captured, however, in other experiments, the number of spectral bands required may be pre-determined by performing principal component analysis. Hence, the time required to capture and process the hyperspectral images can be reduced.

 figure: Fig. 1.

Fig. 1. Schematic drawings of the dual-modality hyperspectral imaging system for (a) transmission modality and (b) fluorescence modality. Note that a beam splitter is not necessary if operating in transmission modality alone.

Download Full Size | PDF

2.2 Sample preparation for transmission imaging

Human breast cancer microscope slides were imaged. They were purchased from PrecisionMed (Accession S09-2325, Block A06). Biopsies of human breast tissues from patients were processed using the standard protocol for formalin-fixed paraffin-embedded tissues. Each tissue sample was sectioned and deparaffinized to prepare for the Hematoxylin and Eosin (H&E) stained pathological slides. H&E stain is the most widely used staining method and is treated as the gold standard. Hematoxylin principally stains the cell nuclei in blue or dark-purple and eosin stains the cytoplasm and extracellular matrix structures such as collagen in pink.

2.2 Sample preparation for fluorescence imaging

Quantum dots (QDots) are widely used fluorophores [14,15]. The quantum dot samples used in this paper were CdSe core, ZnS shell structure (CdSe/ZnS) quantum dots coated with an additional polymer shell and conjugated to streptavidin protein (Q10151MP, Invitrogen). A total of six quantum dot conjugates with different emission spectra were used, with emission peaks at 525 nm (Qdot 525), 565 nm (Qdot 565), 585 nm (Qdot 585), 605 nm (Qdot 605), 655 nm (Qdot 655), and 705 nm (Qdot 705). All samples were stored as 1 µM solutions in 1 M betaine and 50 mM borate solvent with 0.05% sodium azide at pH 8.3. For imaging individual quantum dots, 2 µL of each sample was loaded onto a clean glass slide, while for imaging quantum dot mixtures, two types of mixture samples were prepared. Mixture (I) was a homogeneous mixture of all six quantum dot samples. 0.5 µL of each quantum dot conjugate was added and mixed using a vortex mixer, yielding a total volume of 3 µL of the mixture; out of which, 2 µL of the mixture was used for imaging. Mixture (II) was a heterogeneous mixture containing spatially separated six quantum dot samples, prepared by loading 1 µL of Qdot 525 and Qdot 565 each, and 0.5 µL of the remaining samples onto a clean glass slide without prior mixing. As shown in Fig. 2, samples were loaded into a 3 by 2 matrix format on the glass slide. As samples were loaded close to each other, there were overlapping regions containing heterogeneous mixtures of two different samples; for instance between Qdots 525 and 605, as well as between Qdots 585 and 705. All individual samples and mixture (I) were imaged while still wet as well as after drying, while mixture (II) was imaged only after drying.

 figure: Fig. 2.

Fig. 2. Quantum dot samples for fluorescence imaging include an individual quantum dot sample, a homogeneous mixture sample, and a heterogeneous mixture sample.

Download Full Size | PDF

2.4 Imaging acquisition

For the transmission imaging experiment, an exposure curve was first obtained by recording the exposure time for the wavelength range from 420–730 nm at 11 different wavelengths in a sample-free condition. The desired exposure time was set such that there was no overexposure. The exposure curve may vary for each experimental condition, such as the intensity of the light source, the bandwidth setting of the tunable filter, and the objective used. However, once the exposure curve was set, it could be used for subsequent experiments on the actual samples with the same experimental condition. In the experiment for the H&E stained samples, an exposure curve based on the lamp intensity at 1% output power (40 mW), medium band setting, and 10x objective could be found in the Supplement 1. The exposure time decreased as the wavelength increased because of the transmission characteristics of the liquid crystal tunable filter. A sample-free datacube was first captured as the reference. Image correction was then performed by dividing the sample datacube with the reference datacube to obtain the transmission spectrum of the sample at each pixel.

For the fluorescence imaging experiment with individual quantum dot samples and mixture (I), hyperspectral images were acquired using the 10x objective and the narrow band setting. The exposure duration of 170 ms at 90% LED power (1143 mW) was selected based on the emission of the brightest sample, Qdot 655, such that the hyperspectral image of Qdot 655 would be as bright as possible without any overexposure. Mixture (II) was imaged using the 4x objective to preserve a sufficiently wide field of view during image capture. An exposure time of 900 ms at 90% LED power (1143 mW) and a narrow band setting were selected for optimal exposure of the different regions of interest across the sample, which were the overlapping regions between different quantum dots.

A total of 311 spectral images were collected with wavelength scanning starting from 420 nm to 730 nm at a 1 nm step size. After the acquisition of the hyperspectral images, full-color reconstruction was performed using three single-channel images corresponding to the Red (633 nm), Green (545 nm), and Blue (455 nm) channels. This allows for the hyperspectral image to be quickly visualized and gives an overview of the spatial information captured within the hyperspectral image.

2.5 Spectral angle mapper

Spectral angle mapper (SAM) is a commonly used spectral classification method that uses an n-D angle to match each pixel of a hyperspectral image to a reference spectral library consisting of one or more spectrums of known material classes. The spectral similarity between the sample spectrum and the reference spectrum was determined by calculating the angle between them and treating them as vectors in a space with dimensionality equals to the number of spectral bands. The spectral angle distance can be expressed in the following formula [16]:

$$\alpha = {\cos ^{ - 1}}\left( {\frac{{\mathop \sum \nolimits_{\lambda = 1}^n {t_\lambda }{r_\lambda }}}{{{{(\mathop \sum \nolimits_{\lambda = 1}^n {t_\lambda }^2)}^{0.5}}{{(\mathop \sum \nolimits_{\lambda = 1}^n {r_\lambda }^2)}^{0.5}}}}} \right)$$
where ${t_\lambda }$ and ${r_\lambda }$ are the target and reference intensity values at $\lambda $, and n is the number of spectral bands. SAM compares the angle between the sample vector and the reference vector in the n-D space. A smaller angle represents a closer match to the reference spectrum. A threshold value can be set so that closely matched sample spectra will be grouped and treated as the same material class as the reference spectrum. According to the literature, this method is relatively insensitive to illumination and albedo effects when used on calibrated reflectance data [17].

In this paper, SAM was applied to biomedical imaging results where the spectral angle distance values were used for the identification of different tissue and cell structures on histopathological slides.

2.2 Spectral unmixing

Spectral unmixing is an analysis tool that studies the spectrum of a sample containing multiple substances by decomposing the overall spectrum into the set of its component spectra, or endmembers; as well as their respective proportions, or abundances [18]. For hyperspectral imaging, spectral unmixing can be used in situations where each pixel in a hyperspectral image contains information from a mixture of multiple substances at the sub-pixel scale, or when imaged substances may possess similar spectra and result in high degrees of spectral overlap with each other; in both cases manifesting as a superposition of the spectra of multiple components in the observed spectrum at each image pixel.

While there are many models describing various types of interactions between the different substances at the sub-pixel scale, the linear spectral unmixing model is the simplest and is suitable for describing substances that do not interact with each other. Each photon that reaches the detector is assumed to have only interacted with a single substance with no other optical effects along its pathway to the detector. Mathematically, linear spectral unmixing can be expressed as the following equation:

$$\theta (\lambda )= \mathop \sum \nolimits_\textrm{i}^\textrm{n} {a_i}{\theta _i}(\lambda )+ E$$
where $\theta (\lambda )$ is the superimposed spectrum as a function of the wavelength, ${a_i}$ and ${\theta _i}(\lambda )$ are the abundance and endmember spectrum for the $i$-th endmember respectively, n is the total number of endmembers, and E refers to the observation noise term. Typically, endmember extraction is first performed to determine the set of endmember spectra that comprises the superimposed spectrum. Once the set of endmember spectra (i.e. a spectral library) is known, the abundances of each endmember can then be estimated by the least-squares solution to Eq. (2), typically with non-negative and sum-to-one constraints set upon the estimated abundance values if all endmembers are already known beforehand.

In this paper, linear spectral unmixing was performed on the hyperspectral images of mixtures (I) and (II) for fluorescence imaging. The set of endmember spectra for these images was the spectra of individual quantum dots obtained from the hyperspectral images of the individual samples; as such, endmember extraction does not have to be conducted. The unmixing process was performed on each pixel in the hyperspectral image, taking the superimposed spectrum as well as the known endmember spectra as inputs, and returning the series of calculated abundance values of each endmember as the output. Thus, spectral unmixing on hyperspectral images yields a new data cube of identical spatial dimensions and a reduced number of spectral bands corresponding to the number of endmembers, with each pixel containing the series of unmixed abundance values from its original spectrum.

3. Results and discussion

3.1 Transmission imaging

3.1.1 H&E stained samples

Among all the diagnostic measures, biopsy with pathological examination is the crucial step to validate the presence and severity of cancer. In this section, one breast cancer H&E stained sample was studied as an example. The patient had invasive ductal carcinoma and estrogen and progesterone positive, according to the pathology report. Figure 3(a) is the reconstructed full-color image of a section of the slide. Extracellular matrix, cell nuclei, and air space have been identified based on the color differences. Besides just the color differences, hyperspectral imaging also provides the spectral profiles of each component at their respective pixels. Figure 3(b) shows the spectrum of the air space (dark blue), ECM 1 (orange), ECM 2 (purple), and cell nuclei (yellow). Followed by spectral analysis, a material discovery was performed using the SAM algorithm to show the spatial distribution of different material classes. Figure 3(c) shows the false-color representation of the material discovery results. Regions with similar spectra are classified as one material and assigned a color, so this representation shows which regions of the sample contain the same material class. Yellow represents air space, blue represents ECM 1 and 2, and black represents cell nuclei. In the future, this analysis approach may be useful in the fields of digital staining [19,20].

 figure: Fig. 3.

Fig. 3. Hyperspectral imaging of stained human breast cancer slide (a selected section). (a) Reconstructed full-color image with two extracellular matrix materials, cell nuclei, and air space identified. (b) Extracted spectral information of identified material classes in (a). (c) False-color image based on the material discovery results. Yellow represents air space, blue represents ECM 1 and 2, and black represents cell nuclei.

Download Full Size | PDF

3.1 Fluorescence imaging

Unlike conventional fluorescence imaging, hyperspectral fluorescence imaging allows for the quantification of the relative emission intensities of each quantum dot sample present within each pixel location in the case of a homogeneous mixture (i.e. mixture (I)), and the visualization of the spatial distribution of each quantum dot sample across each image in the case of a heterogeneous mixture (i.e. mixture (II)), by performing spectral unmixing on the spectral information obtained from the hyperspectral images. It is worth mentioning that although the mixtures studied here contain limited types of quantum dots, the same analysis method can be applied to a mixture containing any number of fluorophores of the same type or a mix of different types of fluorophores, as long as their spectral features are different within the 311 spectral bands. Hence, a system with a higher spectral resolution will allow for more fluorophores to be quantified at one location and their individual spatial distribution elucidated.

3.2.1 Individual quantum dot samples

For each sample, we took three hyperspectral images when the sample was wet and one hyperspectral image after the sample had dried. Figure 4(a) shows the full-color images reconstructed from the hyperspectral images of the edges of the six quantum dot samples when the samples were dry, with the “coffee-ring” effect visible in these images. Figure 4(b) shows the emission spectrum of each sample, which was generated from the hyperspectral images of the wet samples (see Supplement 1) since the samples were more uniform in the wet condition. The average emission spectrum was calculated across the entire field of view for each hyperspectral image. The obtained spectra closely matched with manufacturer-provided data in terms of the shape, position, and width of the emission peaks; thus demonstrating that our hyperspectral fluorescence microscope system is able to capture the spectral information of imaged samples with a high degree of accuracy. The obtained spectra were subsequently used to build the spectral library for the spectral unmixing analysis of mixture (I) and mixture (II). It should be noted that the average spectra of the individual quantum dot samples had different peak intensities, due to differences in the extinction coefficients of each quantum dot sample for the same excitation wavelength and intensity. Quantum dots are widely used to develop tissue-mimicking phantoms for fluorescence imaging. Based on the literature on the excitation and emission wavelengths of clinically used fluorescent agents in the visible range [21], Qdot 525 has similar properties as FITC (EC-17, etc.), HMRG (GCP-001, etc.), fluorescein sodium, and BODIPY FL (PARPi-FL); Qdot 605 and Qdot 655 have similar properties as 5-ALA.

 figure: Fig. 4.

Fig. 4. (a) Reconstructed full-color images of individual quantum dot samples (edge of sample, 10x magnification) after samples had dried. Images on display were captured using a larger area of the sensor array (3200 × 2500 pixels) and with a larger spectral step size of 5 nm. Note that these were not used for building the spectral library. (b) Normalized average emission spectra of individual quantum dot samples with peak wavelengths and FWHM values are shown. *The exact FWHM value of the Qdot 705 sample could not be determined as the full emission profile was out of the spectral scanning range of the tunable filter; hence an estimated value is shown assuming that the spectral profile is symmetrical.

Download Full Size | PDF

Comparing the hyperspectral images of the same region of the same sample in the wet and dry conditions, as shown in Fig. 5(a) and 5(b), respectively. Interestingly, we observed that there was a spectral redshift of around 5 nm, as shown in Fig. 5(c). This spectral shift was observed in all of the samples (see Supplement 1), which is likely due to the aggregation of the quantum dot particles during the drying process over a period of time (approximately 15 minutes in our case) [22]. We note that this has not been discussed before in the literature and it may have a wide implication for nanoparticles-based fluorescence imaging.

 figure: Fig. 5.

Fig. 5. (a) Reconstructed full-color image of Qdot 565 sample when wet and (b) dry. (c) Normalized average emission spectra from the hyperspectral images of (a) and (b), with a peak intensity shift of around 5 nm.

Download Full Size | PDF

3.2.2 Mixture (I)—quantification of the relative proportions of each quantum dot

Mixture (I) is a homogeneous mixture of equal proportions of all six quantum dot samples. Linear spectral unmixing was performed on the emission spectra of mixture (I) with six endmembers; each endmember corresponding to the emission spectrum of each quantum dot sample as obtained in the previous section. This allows us to calculate their relative proportions in the mixture solution.

Figure 6 shows the reconstructed full-color images of the same region of mixture (I) while both wet and dry, as well as the average spectra and spectrally unmixed proportions of each type of quantum dot sample within the hyperspectral images. Figure 6(a) shows an overall uniform distribution, while Fig. 6(d) shows patches of uneven quantum dot aggregates. Figure 6(b) and 6(e) show the emission spectra of mixture (I) in the wet and dry conditions, respectively. The solid lines are the observed superimposed spectrum of the mixture and the dotted lines are the decomposed spectra of the corresponding endmembers. Both the wet and dry samples appeared to have two prominent peaks at approximately 595 nm and 670 nm, with the fluorescence emissions of Qdot 585 and Qdot 605 contributing to the peak at 595 nm and the fluorescence emissions of Qdot 655 and Qdot 705 contributing to the peak at 670 nm. As the fluorescence emissions of Qdot 525 and Qdot 565 were much weaker compared to the rest of the quantum dot samples for the same excitation condition, their contributions were much less significant. Unlike the individual quantum dot samples, there was a noticeable change in the shape of the spectra between the wet and dry samples of mixture (I) which explains the differences in color of their respective full-color images; with the 595 nm peak having a higher intensity than the 670 nm peak in the wet sample, while the 670 nm peak having a higher intensity than the 595 nm peak in the dry sample. Two possible factors that could contribute to the spectral shift between the wet and dry samples are the aggregation of quantum dot particles as mentioned previously, as well as the inter-quantum dot energy transfer from the quantum dots with lower peak wavelength to those with higher peak wavelength, thus quenching the emission intensity of the quantum dots with a lower peak wavelength while strengthening the emission intensity of the quantum dots with a higher peak wavelength [23]. This inter-quantum dot energy transfer is reflected in the differences between the unmixed proportions of the wet and the dry sample, with the dry sample displaying a general linearly increasing trend with the peak wavelength of the quantum dot samples. Figure 6(c) and (f) are the expected and calculated (spectrally unmixed) relative proportions of each quantum dot sample in the mixture in the wet and dry conditions, respectively. The expected proportions are shown as dashed lines and the calculated or unmixed proportions are shown as bar graphs. The relative proportions were calculated based on the decomposed peak intensities of each quantum dot sample since the peak intensity is proportional to the concentration of the sample. The calculated proportions are relatively close to the expected values with the wet sample matching better than the dry sample (see Supplement 1). One source of error between the calculated proportions and the expected values is due to the interactions between different quantum dot samples, which suggests that a nonlinear spectral unmixing may be a more accurate representation of the mixture sample [24]. Nonetheless, the linear spectral unmixing model adopted here is simple enough to be used for estimated quantitative analysis. The other source of error could be the pipetting error and mixing error because of the small sample volume.

 figure: Fig. 6.

Fig. 6. (a) Reconstructed full-color image of the mixture (I) of all six quantum dot samples (center of the sample, 10x magnification) in the wet condition. (b) Measured emission spectrum (solid line) and decomposed spectra (dotted lines) of (a). (c) Spectrally unmixed proportions of each type of quantum dot sample of (a). (d) A reconstructed full-color image of the mixture (I) in the dry condition. (e) Measured emission spectrum and decomposed spectra of (d). (f) Spectrally unmixed proportions of each type of quantum dot sample of (d). In (c) and (f), the error bars represent the standard deviations from the calculations at three pixel locations.

Download Full Size | PDF

3.2.3 Mixture (II)—quantification of the spatial distribution of each quantum dot

Mixture (II) is a heterogeneous mixture where two types of quantum dots were spatially overlapped, as shown in Fig. 1. For mixture (I), we focused on the spectral features of the hyperspectral images. Hence, for mixture (II), we demonstrate our system’s ability to spatially resolve different fluorophores in a mixture.

Figure 7(a) shows the reconstructed full-color image of the mixture sample. Subsequently, spectral unmixing was performed on the hyperspectral images to spatially distinguish the different regions in each image from each other based on the constituent quantum dot sample. Since spectral information is saved at each pixel, an image can be processed to show regions where the spectral features are the same or similar. For each imaged region, only the quantum dot samples present within the region were included in the unmixing process as endmembers. Figure 7(b) and 7(c) show the spatial distribution of each quantum dot sample based on the spectral unmixing results as a grayscale image. They represent the abundance values in Eq. (2) at each pixel of an endmember. Note that the spatial distributions obtained using this method are more accurate than those obtained based on the single-band images at the peak emission wavelength of each quantum dot sample because spectral unmixing takes into account the whole spectral information and the spectral overlap between different quantum dots in each spectral band.

 figure: Fig. 7.

Fig. 7. (a) Reconstructed full-color image of the region of overlap between Qdot 585 and Qdot 705 (4x magnification). (b) Spectrally unmixed grayscale images displaying the spatial distributions of Qdot 585 and (c) Qdot 705.

Download Full Size | PDF

4. Conclusion

We have presented a hyperspectral imaging microscopy system that is able to observe the transmission and fluorescent properties of the samples and perform dual-modality imaging. Hyperspectral imaging saves data in three dimensions in the spatial and spectral domains, thus providing extra information about the sample than a conventional white light microscopy system and a conventional fluorescent microscopy system. For transmission imaging applications, we have demonstrated that hyperspectral imaging allows the identification of different tissue and cell structures of human histopathology slides using the SAM analysis method. For fluorescence imaging applications, we have demonstrated that hyperspectral imaging is well suited for multiplexed fluorescence imaging where multiple fluorophores are present in a mixture environment. Specifically, we have shown that by using the spectral unmixing method, the relative proportions of each quantum dot fluorophore at one location can be determined and the spatial distribution of each quantum dot type can also be elucidated with minimal crosstalk. A great application is for imaging the resected tissues in fluorescence-guided surgery, especially when two different exogenous fluorescence tracers are used, for example [25], where individual fluorescence tracers can be analyzed in a mixture environment. Our results provide valuable information for the fields of digital staining and digital pathology as well as in vivo fluorescence imaging with multiple fluorophores.

Funding

Agency for Science, Technology and Research (BMRC Central Research Fund (CRF, UIBR) Award, Career Development Award (202D800042)).

Acknowledgments

The authors would like to acknowledge the funding support from A*STAR Career Development Award (202D800042) and BMRC Central Research Fund (CRF, UIBR) Award. The authors would also like to thank Hongbin Xie from Thorlabs for his generous support in instrument development.

S. Z. conceived the experiments. J. Y. C. and J. J. C conducted the experiments. S. Z., J. Y. C., and J. J. C. analyzed the results. X. L. discussed the results. M. O. supervised the project. All authors reviewed the manuscript.

Disclosures

The authors declare no conflicts of interest.

Data Availability

Raw data is available upon request.

Supplemental document

See Supplement 1 for supporting content.

References

1. G. Lu and B. Fei, “Medical hyperspectral imaging: a review,” J. Biomed. Opt. 19(1), 010901 (2014). [CrossRef]  

2. Q. Li, X. He, Y. Wang, H. Liu, D. Xu, and F. Guo, “Review of spectral imaging technology in biomedical engineering: achievements and challenges,” J. Biomed. Opt. 18(10), 100901 (2013). [CrossRef]  

3. N. Mehta, S. Shaik, R. Devireddy, and M. R. Gartia, “Single-cell analysis using hyperspectral imaging modalities,” J. Biomech. Eng. 140(2), 0208021 (2018). [CrossRef]  

4. J. D. Pallua, A. Brunner, B. Zelger, C. W. Huck, M. Schirmer, J. Laimer, D. Putzer, M. Thaler, and B. Zelger, “New perspectives of hyperspectral imaging for clinical research,” NIR news 32(3-4), 5–13 (2021). [CrossRef]  

5. J. Yoon, “Hyperspectral imaging for clinical applications,” BioChip J. 16(1), 1–12 (2022). [CrossRef]  

6. S. Ortega, M. Halicek, H. Fabelo, R. Camacho, M. Plaza De La Luz, F. Godtliebsen, G. M. Callicó, and B. Fei, “Hyperspectral imaging for the detection of glioblastoma tumor cells in h&e slides using convolutional neural networks,” Sensors 20(7), 1911 (2020). [CrossRef]  

7. Y. Khouj, J. Dawson, J. Coad, and L. Vona-Davis, “Hyperspectral imaging and K-means classification for histologic evaluation of ductal carcinoma in situ,” Front. Oncol. 8, 17 (2018). [CrossRef]  

8. S. Ortega, H. Fabelo, R. Camacho, M. de la Luz Plaza, G. M. Callicó, and R. Sarmiento, “Detecting brain tumor in pathological slides using hyperspectral imaging,” Biomed. Opt. Express 9(2), 818 (2018). [CrossRef]  

9. Y.-D. Kim, D. Do, H. Yoo, and D. Gweon, “Multimodal confocal hyperspectral imaging microscopy with wavelength sweeping source,” Meas. Sci. Technol. 26(2), 025701 (2015). [CrossRef]  

10. W. Shi, D. E. S. Koo, M. Kitano, H. J. Chiang, L. A. Trinh, G. Turcatel, B. Steventon, C. Arnesano, D. Warburton, S. E. Fraser, and F. Cutrale, “Pre-processing visualization of hyperspectral fluorescent data with spectrally encoded enhanced representations,” Nat. Commun. 11(1), 726 (2020). [CrossRef]  

11. A. D. Elliott, L. Gao, A. Ustione, N. Bedard, R. Kester, D. W. Piston, and T. S. Tkaczyk, “Real-time hyperspectral fluorescence imaging of pancreatic β-cell dynamics with the image mapping spectrometer,” J. Cell Sci. 125(1), 121–132 (2012). [CrossRef]  

12. R. Fakhrullin, L. Nigamatzyanova, and G. Fakhrullina, “Dark-field/hyperspectral microscopy for detecting nanoscale particles in environmental nanotoxicology research,” Sci. Total Environ. 772, 145478 (2021). [CrossRef]  

13. P. Zamora-Perez, D. Tsoutsi, R. Xu, and P. Rivera Gil, “Hyperspectral-enhanced dark field microscopy for single and collective nanoparticle characterization in biological environments,” Materials 11(2), 243 (2018). [CrossRef]  

14. F. P. García de Arquer, D. V. Talapin, V. I. Klimov, Y. Arakawa, M. Bayer, and E. H. Sargent, “Semiconductor quantum dots: technological progress and future challenges,” Science 373(6555), eaaz8541 (2021). [CrossRef]  

15. A. Foubert, N. V. Beloglazova, A. Rajkovic, B. Sas, A. Madder, I. Y. Goryacheva, and S. De Saeger, “Bioconjugation of quantum dots: review & impact on future application,” TrAC Trends Anal. Chem. 83, 31–48 (2016). [CrossRef]  

16. F. A. Kruse, A. B. Lefkoff, J. W. Boardman, K. B. Heidebrecht, A. T. Shapiro, P. J. Barloon, and A. F. H. Goetz, “The spectral image processing system (SIPS)—interactive visualization and analysis of imaging spectrometer data,” Remote Sens. Environ. 44(2-3), 145–163 (1993). [CrossRef]  

17. X. Jia, “Spectral angle mapper,” in Field Guide to Hyperspectral/Multispectral Image Processing (SPIE, 2022), pp. 345–346.

18. N. Keshava and J. F. Mustard, “Spectral unmixing,” IEEE Signal Process. Mag. 19(1), 44–57 (2002). [CrossRef]  

19. S. Ortega, H. Fabelo, D. Iakovidis, A. Koulaouzidis, and G. Callico, “Use of hyperspectral/multispectral imaging in gastroenterology. shedding some–different–light into the dark,” J. Clin. Med. 8(1), 36 (2019). [CrossRef]  

20. Y. Zhang, K. de Haan, Y. Rivenson, J. Li, A. Delis, and A. Ozcan, “Digital synthesis of histological stains using micro-structured and multiplexed virtual staining of label-free tissue,” Light: Sci. Appl. 9(1), 78 (2020). [CrossRef]  

21. L. J. Lauwerends, H. Abbasi, T. C. Bakker Schut, P. B. A. A. Van Driel, J. A. U. Hardillo, I. P. Santos, E. M. Barroso, S. Koljenović, A. L. Vahrmeijer, R. J. Baatenburg de Jong, G. J. Puppels, and S. Keereweer, “The complementary value of intraoperative fluorescence imaging and Raman spectroscopy for cancer surgery: combining the incompatibles,” Eur. J. Nucl. Med. Mol. Imaging 49(7), 2364–2376 (2022). [CrossRef]  

22. V. Biju, Y. Makita, A. Sonoda, H. Yokoyama, Y. Baba, and M. Ishikawa, “Temperature-sensitive photoluminescence of CdSe quantum dot clusters,” J. Phys. Chem. B 109(29), 13899–13905 (2005). [CrossRef]  

23. R. Kanemoto, A. Anas, Y. Matsumoto, R. Ueji, T. Itoh, Y. Baba, S. Nakanishi, M. Ishikawa, and V. Biju, “Relations between dewetting of polymer thin films and phase-separation of encompassed quantum dots,” J. Phys. Chem. C 112(22), 8184–8191 (2008). [CrossRef]  

24. N. Dobigeon, Y. Altmann, N. Brun, and S. Moussaoui, “Linear and nonlinear unmixing in hyperspectral imaging,” Data Handl. Sci. Technol. 30, 185–224 (2016). [CrossRef]  

25. K. M. Tichauer, K. S. Samkoe, J. R. Gunn, S. C. Kanick, P. J. Hoopes, R. J. Barth, P. A. Kaufman, T. Hasan, and B. W. Pogue, “Microscopic lymph node tumor burden quantified by macroscopic dual-tracer molecular imaging,” Nat. Med. 20(11), 1348–1353 (2014). [CrossRef]  

Supplementary Material (1)

NameDescription
Supplement 1       Supplementary information (revised)

Data Availability

Raw data is available upon request.

Cited By

Optica participates in Crossref's Cited-By Linking service. Citing articles from Optica Publishing Group journals and other participating publishers are listed here.

Alert me when this article is cited.


Figures (7)

Fig. 1.
Fig. 1. Schematic drawings of the dual-modality hyperspectral imaging system for (a) transmission modality and (b) fluorescence modality. Note that a beam splitter is not necessary if operating in transmission modality alone.
Fig. 2.
Fig. 2. Quantum dot samples for fluorescence imaging include an individual quantum dot sample, a homogeneous mixture sample, and a heterogeneous mixture sample.
Fig. 3.
Fig. 3. Hyperspectral imaging of stained human breast cancer slide (a selected section). (a) Reconstructed full-color image with two extracellular matrix materials, cell nuclei, and air space identified. (b) Extracted spectral information of identified material classes in (a). (c) False-color image based on the material discovery results. Yellow represents air space, blue represents ECM 1 and 2, and black represents cell nuclei.
Fig. 4.
Fig. 4. (a) Reconstructed full-color images of individual quantum dot samples (edge of sample, 10x magnification) after samples had dried. Images on display were captured using a larger area of the sensor array (3200 × 2500 pixels) and with a larger spectral step size of 5 nm. Note that these were not used for building the spectral library. (b) Normalized average emission spectra of individual quantum dot samples with peak wavelengths and FWHM values are shown. *The exact FWHM value of the Qdot 705 sample could not be determined as the full emission profile was out of the spectral scanning range of the tunable filter; hence an estimated value is shown assuming that the spectral profile is symmetrical.
Fig. 5.
Fig. 5. (a) Reconstructed full-color image of Qdot 565 sample when wet and (b) dry. (c) Normalized average emission spectra from the hyperspectral images of (a) and (b), with a peak intensity shift of around 5 nm.
Fig. 6.
Fig. 6. (a) Reconstructed full-color image of the mixture (I) of all six quantum dot samples (center of the sample, 10x magnification) in the wet condition. (b) Measured emission spectrum (solid line) and decomposed spectra (dotted lines) of (a). (c) Spectrally unmixed proportions of each type of quantum dot sample of (a). (d) A reconstructed full-color image of the mixture (I) in the dry condition. (e) Measured emission spectrum and decomposed spectra of (d). (f) Spectrally unmixed proportions of each type of quantum dot sample of (d). In (c) and (f), the error bars represent the standard deviations from the calculations at three pixel locations.
Fig. 7.
Fig. 7. (a) Reconstructed full-color image of the region of overlap between Qdot 585 and Qdot 705 (4x magnification). (b) Spectrally unmixed grayscale images displaying the spatial distributions of Qdot 585 and (c) Qdot 705.

Equations (2)

Equations on this page are rendered with MathJax. Learn more.

α = cos 1 ( λ = 1 n t λ r λ ( λ = 1 n t λ 2 ) 0.5 ( λ = 1 n r λ 2 ) 0.5 )
θ ( λ ) = i n a i θ i ( λ ) + E
Select as filters


Select Topics Cancel
© Copyright 2024 | Optica Publishing Group. All rights reserved, including rights for text and data mining and training of artificial technologies or similar technologies.