Open Access
Add to BibSonomy
Abstract
We propose a confocal hyperspectral microscopic imager (CHMI) that can measure both transmission and fluorescent spectra of individual microalgae, as well as obtain classical transmission images and corresponding fluorescent hyperspectral images with a high signal-to-noise ratio. Thus, the system can realize precise identification, classification, and location of microalgae in a free or symbiosis state. The CHMI works in a staring state, with two imaging modes, a confocal fluorescence hyperspectral imaging (CFHI) mode and a transmission hyperspectral imaging (THI) mode. The imaging modes share the main light path, and thus obtained fluorescence and transmission hyperspectral images have point-to-point correspondence. In the CFHI mode, a confocal technology to eliminate image blurring caused by interference of axial points is included. The CHMI has excellent performance with spectral and spatial resolutions of 3 nm and 2 µm, respectively (using a 10× microscope objective magnification). To demonstrate the capacity and versatility of the CHMI, we report on demonstration experiments on four species of microalgae in free form as well as three species of jellyfish with symbiotic microalgae. In the microalgae species classification experiments, transmission and fluorescence spectra collected by the CHMI were preprocessed using principal component analysis (PCA), and a support vector machine (SVM) model or deep learning was then used for classification. The accuracy of the SVM model and deep learning method to distinguish one species of individual microalgae from another was found to be 96.25% and 98.34%, respectively. Also, the ability of the CHMI to analyze the concentration, species, and distribution differences of symbiotic microalgae in symbionts is furthermore demonstrated.
© 2021 Optical Society of America under the terms of the OSA Open Access Publishing Agreement
1. Introduction
Microalgae are indispensable parts of marine ecosystems, playing important roles in photosynthesis, energy transformation, and are furthermore useful for water quality evaluation [1]. They form symbionts with many marine organisms, and provide nutrients, however microalgae outbreak is also a major contributor to marine pollution, which is a danger to other marine organisms, can cause significant damage to aquaculture and pose a safety threat to hydroelectric power plants. Therefore, development of efficient microalgae detection and monitoring system is highly pertinent, not only for the purpose of monitoring growth and health status of marine organisms symbiotic with microalgae, but also for early warning of microalgae outbreaks [2]. Conventional microalgae detection is typically based on microscopic observation that require highly trained professionals for identification and classification based on morphological features, which is costly and time consuming [3,4].
Hyperspectral imaging can acquire both spatial and spectral information, which enables non-contact and efficient detection with high throughput. Hyperspectral imaging provides a rich set of information related to vibrational modes of molecular bonds and light absorption, which can be used in material identification and classification. At present, it has a broad range of application areas, e.g., online environmental monitoring, food safety control, soil classification, gas monitoring, biomedicine, etc [5–14]. Hyperspectral imaging includes fluorescence hyperspectral imaging and transmission hyperspectral imaging, etc. Average transmission spectra have been used to distinguished different species of microalgae as the transmission spectra reflect the optical absorption characteristics of pigment substances in the different microalgae species [2,15]. However, this works poorly when light absorption of the pigment substances is low, which results in poor sensitivity for detection of individual microalgae. For this reason, it is beneficial to, in addition, combine and utilize other optical features for microalgae species classification. The sensitivity of fluorescence spectrum is 2–3 orders of magnitude higher than that of absorption spectrum and the chlorophyll in microalgae can be excited to emit fluorescence. As the chlorophyll contents in varying species of microalgae are different, the fluorescence spectra can be used to identify the microalgae species as well [16].
However, conventional fluorescence hyperspectral imagers suffer from low resolution and blurry image quality, which is caused by interference of fluorescent signals along the same axis. Such fluorescence interference can be eliminated by using confocal laser scanning microscope (CLSM) technique, which provides higher signal-to-noise ratio (SNR) and imaging contrast compared with a conventional light microscope. The CLSM principle of operation is as follows: an excitation laser is focused on a point of the sample and excites fluorescence light. After passing through a series of optical components, the fluorescent light converges on a pinhole and is received by a photodetector. In this way, only the fluorescence generated at the point where the excitation light is focused is detected, while fluorescence emitted from other axial positions is blocked by the pinhole. Combining such a CLSM module with a hyperspectral imager enables precise spectral detection and analysis of a single fluorescence point [17–20].
Few experimental studies combing fluorescence and transmission hyperspectral to analyze optical characteristics of measured samples have been reported to date. Taniguchi et al. recently compiled the absorption and fluorescence Spectral Database of Chlorophylls and Analogues, which facilitates comparisons and quantitative calculations [21]. Nozue et al. used spectroscopic microscopy to analyze fluorescence and absorption spectra of thylakoid membranes in Rivularia and variabilis cells [22]. Baszanowska et al. utilized spectral signatures of fluorescence and light absorption to identify crude oils in a marine environment [23]. However, the setups used in these studies cannot provide point-to-point correspondence between the transmission and fluorescence spectral information of a single pixel, which limits their applications in substance identification and classification.
In this paper, we propose a confocal hyperspectral microscopic imager (CHMI) that enables precise identification, classification, and location of microalgae in free or symbiosis states. The CHMI integrates a confocal fluorescence hyperspectral imaging (CFHI) mode and a transmission hyperspectral imaging (THI) mode. For the CFHI mode, we introduce CLSM to eliminate image blurring that is otherwise caused by interference fluorescence from different axial points. Thus, the spectrum of a single fluorescence point and fluorescent hyperspectral images with high signal-to-noise ratio can be obtained. Since both THI and CFHI modes share a major light path, point-to-point correspondence can be achieved between the fluorescence and transmission hyperspectral images. Transmission spectra provide information of the light absorption characteristics of e.g., chlorophyll, carotenoid, while the fluorescence spectra provide luminescent characteristics of fluorescent substances (being chlorophyll in this study). The combined spectral information suffices to differentiate between microalgae species in micro size as well as to characterize these. The fluorescence images furthermore clearly show the distribution of microalgae and the location of symbiotic microalgae in symbionts. The CHMI has excellent performance with a spectral resolution of 3 nm and spatial resolution of 2 µm (for 10× microscope objective magnification). To demonstrate the CHMI’s performance, it was used to identify and classify four species of microalgae in low concentrations. Transmission spectra combined with fluorescence spectra collected by the CHMI were preprocessed using principal component analysis (PCA) [24], and a support vector machine (SVM) [25–27] model or deep learning was used for classification. The classification accuracy of the two analysis methods were found to be 96.25% and 98.34%, respectively, indicating that integrating CFHI and THI modes significantly improve the classification accuracy for microbiological specimens. We also report on characterizations of three species of jellyfish with symbiotics that furthermore demonstrate the capabilities of the CHMI to analyze concentration, species, and distribution differences of symbiotic microalgae in symbionts. The results demonstrate that the hyperspectral imager that we propose has good spectral detection performance of micro-sized biological samples. The CHMI greatly improves accuracy of identification, classification, and location of microalgae in low concentrations.
The paper is structured as follows: Section II introduces the CHMI and describes its principle of operation as well as the wavelength calibration of the hyperspectral imager. In Section III, several experiments that demonstrate the CHMI performance are described, including hyperspectral imaging experiments of four species of microalgae based on the CHMI combined with SVM or deep learning to classify microalgae species, and characterizations of three species of jellyfish with symbiotic microalgae. The paper ends with a summary and outlook.
2. Methods
2.1 System setup and calibration
The CHMI can operate in two different imaging modes: the CFHI and THI modes. As mentioned in the introduction, we introduced a confocal technique in the CFHI mode to eliminate mutual interference of excited fluorescence. In this way, the fluorescence spectrum of single point can be analyzed, and fluorescence images with high resolution can be obtained. As the two imaging modes share a same main optical path, there exist a point-to-point correspondence between the transmission and fluorescence hyperspectral images, i.e., the CHMI is capable of transmission as well as fluorescence spectrum analysis of a single point. Combining these two kinds of spectra, the CHMI can characterize the optical properties of samples based on two different types of spectral characteristics, which can improve the accuracy of material detection and classification, especially so for micro-sized objects.
Figure 1 shows the CHMI setup in CFHI mode and the operating principle to obtain fluorescence hyperspectral images is as follows: The excitation light source is a 532 nm laser, the light from which is collimated and launched into an optical fiber. Some parameters of the 532 nm laser are given as follow. Its spot quality factor M2 is less than 1.1, and the spot mode is TEM00. The operation model is continue-wave, and the output power is adjustable in the range of 0∼100 mW. The light is introduced into the system through a reflector and a beam splitter (BS-1) and enters a galvanometer, which can adjust the angles of the incident light such that the excitation light can irradiate different points of the sample. A scan lens and a telecentric lens (ITL200) are used to optimize the scanning image plane and increase the spot diameter. In this way, excitation light at different angles can always converge onto the entry pupil of the microscope objective lens after passing through a reflector. By adjusting the distance between the microscope objective and the sample using a motorized positioning system, the sample can be placed in the focal plane of the microscope. Thus, the excitation light can always be focused on the sample. The point of the sample that is irradiated by the 532 nm laser generates fluorescent light, which, after being collimated by the microscope objective, will return along the original optical path sequentially passing through the reflector, telecentric lens, scan lens, galvanometer and BS-1. A 532 nm long-pass filter is used to filter out the excitation light, and thus the excitation light at 532 nm and the fluorescent light of the sample can be separated. The fluorescent light is then split by a 50/50 BS-2, with one path converging at the pinhole position through an aspheric achromatic lens (AAL-1). With this setup only fluorescent light generated at the point of the sample where the excitation light is focused at can pass through the pinhole, whereas fluorescent light generated elsewhere is blocked. In this way, the interference of axial or adjacent fluorescent light is eliminated, and the single fluorescence point is clear. Fluorescent light from the other path of BS-2 is collected by the hyperspectral imager and, similar to the pinhole, the slit of the hyperspectral imager also eliminates interference from other fluorescence points. By varying the angles of the excitation light using the galvanometer, fluorescent light excited at different points of the sample will pass through the pinhole sequentially and form a complete fluorescence image, while at the same time the corresponding fluorescence spectra is collected by the hyperspectral imager. The fluorescence spectra and fluorescence image together form the fluorescence hyperspectral image.
Figure 2 shows the CHMI setup in THI mode for which the operating principle is as follows: A white light LED is used as the light source that irradiates the sample through a reflector. After the sample is imaged by a microscope objective, a relay image plane is formed at the position of the common focal plane of the ITL200 and the scan lens. This relay image plane is conjugate with the transmission hyperspectral image plane and confocal fluorescence image plane, and can be used to observe the position of the sample and determine whether it is focused or not. The light from the relay image plane is collimated by the scan lens and passed through the galvanometer, BS-1, BS-2 and an AAL-2, after which it reaches the hyperspectral imager. After passing through the slit of the hyperspectral imager, the spectral image of a line region can be collected. Varying the angles of the incident light using the galvanometer, different line regions of the image can pass through the slit in sequence to form corresponding spectral images. The spectral images of the line regions are then spliced together to form a complete transmission hyperspectral image.
Both the fluorescence and transmission spectra are produced by a staring hyperspectral imager, depicted in Figs. 1 and 2. The hyperspectral imager consists of two AALs, a slit, a prism-grating-prism (PGP) pair, and a monochrome imaging CMOS (ZWO, ASI174). The spot incident light generated in CFHI mode or line incident light generated in THI mode reaches the hyperspectral imager and first passes through the slit, after which the light is collimated by an AAL-3 reaching the PGP pair. It is dispersed and expanded in the spectral range 450–800 nm by the PGP, and then focused on the CMOS by the AAL-4. In this way, the spectrum of one point or line region of the image is be obtained. As the galvanometer changing the angles of the excitation laser or incident light, different points or line regions of the image can pass through the slit continuously and finally form a complete fluorescence or transmission hyperspectral image.
The full CHMI combines the CFHI and THI modes to obtain both fluorescence and transmission hyperspectral images of the sample. As the two imaging modes share a main optical path, there exist point-to-point correspondence between the two hyperspectral images, and thus transmission and fluorescence spectrum analysis of a single point can be done. Based on the optical model described in the above, we designed and built up the CHMI in-house, shown in Fig. 3. Designing and building the full system in-house independent of a commercial microscope system allows for a flexible system that can easily be extended.
2.2 Wavelength calibration and resolution measurement of the CHMI
Prior to hyperspectral imaging, wavelength calibration of the spectral images is required to enable the transformation of pixel index of the spectral axis to wavelengths. A Mercury Argon Calibration Source, which emits multiple extremely narrow spectral lines, irradiates the slit of the hyperspectral imager. After diffraction by the PGP structure, a corresponding spectral image of the calibration light source is collected by the CMOS, as shown in Fig. 4(a). Five distinct spectral lines located at the wavelengths, λ, 435.8 nm, 546.1 nm, 578.0 nm, 696.543 nm, and 750.4 nm, respectively, which corresponds to the characteristic spectral peaks of the light source, are chosen. The corresponding pixel indices, y, are 253, 563, 648, 962 and 1107. The relationship between the wavelength index and the pixel coordinates can be described by a following polynomial equation [28]:
where λ is the calibrated wavelength vector, y is the pixel index vector, and a0, a1, a2, and a3 are calibration coefficients. Using a polynomial least square method, the calibration coefficients are calculated to be [a0, a1, a2, a3] = [356.5038, 0.2877, 1.1397×10−4, −4.7396×10−8]. The fitted curve of the polynomial equation is shown in Fig. 4(b) in the spectral range 400–800 nm. As seen in Fig. 4(c), the full width at half maximum (FWHM) of the spectral line 546.1 nm is about 3 nm, i.e., the spectral resolution of the CHMI is about 3 nm.
Fig. 4. (a) The spectral image of the Mercury Argon Calibration Source. (b) The relationship between the wavelength and pixel index fitted by third-order polynomial. (c) The spectral curve of the Mercury Argon Calibration Source measured by our hyperspectral imager after wavelength calibration.
The hyperspectral images captured by the CHMI have a high signal-to-noise ratio. To test the spatial resolution, we performed an imaging experiment in both CFHI and THI modes using a standard resolution test board that has an ultimate resolution of 2 µm. A 10× magnification microscope objective (OLYMPUS Inc., LUMPLFLW series) was used. Figure 5(a) shows the board in the relay image plane, which is clear and can be used for preliminarily observation and selection the sample imaging area. Figure 5(b) shows the image result in CFHI mode, where the minimum resolution patterns can be clearly seen, indicating that the spatial resolution in CFHI mode is less than 2 µm. Similarly, Fig. 5(c) and 5(d) show the image result in THI mode and its magnified portion, indicating that the spatial resolution of THI mode is less than 2 µm as well. The SNR of an image can be approximately equal to the ratio of signal variance to noise variance [29]. Images with SNR higher than 30 dB are considered as images with good SNR. The SNR of fluorescence hyperspectral image in Fig. 5(b) is 58 dB, and that of transmission hyperspectral image in Fig. 5(c) is 41 dB.
Fig. 5. (a) Relay image of the standard resolution test board. (b) Image of the standard resolution test board captured in CFHI mode. (c) Image of the standard resolution test board captured in THI mode. (d) Enlargement of the red rectangular region in Fig. 5(c).
3. Experimental demonstrations
In this section, we describe experiments done using the CHMI to demonstrate its performance. We used it combined with computational methods to classify four different species of microalgae in low concentrations. We furthermore demonstrate the ability of the CHMI to monitor symbionts in an experiment on three species of jellyfish.
3.1 Hyperspectral imaging experiments of four species of microalgae
Microalgae play an important role in photosynthesis and energy transformation and thus classification, identification, and monitoring of microalgae are of significance. To demonstrate the performance of the combined CFHI and THI modes of the CHMI, a series of hyperspectral imaging detections were conducted. Microscopic images of four species of microalgae used as the samples, chlamydomonas, phaeocystis, chlorella, chaetoceros, are shown in Figs. 6(a), 6(b), 6(c) and 6(d). Here, we utilized a 40×magnification microscope objective (OLYMPUS Inc., LUMPLFLW series) to observe the microalgae. The four species of microalgae were supplied by South China Fisher Research Institute (CAFS), and cultured in sea water. In the experiment, we pipetted the microalgae solution into a breaker and diluted the solution. The diluted solution was shaken and dropped onto a glass slide for detection. The microalgae solution concentration is about 1.66×105 cell mL−1. The microalgae are different in color, size, and chlorophyll content. The concentrations of these four species are relatively low, and the microalgae are mainly in free form. Figures 6(e), 6(f), 6(g) and 6(h) show the transmission hyperspectral images of four species of microalgae mentioned above in which their shapes and sizes can be easily differentiated, demonstrating the high-resolution hyperspectral imaging of our system. In the CFHI mode, the chlorophyll of the microalgae is excited by a 532 nm laser and emits fluorescent light, which is captured by the hyperspectral imager. The fluorescence hyperspectral images are shown in Figs. 6(i), 6(j), 6(k) and 6(l). As the CLSM eliminates interference as discussed above, the space between different microalgae as well as their edge profiles are clear. Thus, the fluorescence imaging can be used for map the distribution of microalgae.
Fig. 6. Microscopic images of (a) chlamydomonas, (b) phaeocystis, (c) chlorella and (d) chaetoceros; transmission hyperspectral images of (e) chlamydomonas, (f) phaeocystis, (g) chlorella and (h) chaetoceros; and fluorescence hyperspectral images of (i) chlamydomonas, (j) phaeocystis, (k) chlorella and (l) chaetoceros. Scale bar: 50 µm.
The normalized transmission and fluorescence intensity spectra are shown in Figs. 7(a) and 7(d), respectively. The spectrum of LED source is also shown in Fig. 7(a) for reference. The intensity of the spectrum curves of the microalgae are lower than that of light source at 450 nm due to the optical absorption of chlorophyll. To eliminate the influence of the LED light source and better analyze the inherent optical properties of the samples, we derive normalized transmissivity spectra by dividing the intensity spectra of the microalgae by the LED source intensity spectrum, shown in Fig. 7(b). The results show that the four species of microalgae have strong absorption at 450 nm and 670 nm, which corresponds to the optical absorption of chlorophyll. At 670 nm, chlamydomonas has the highest optical absorption, followed by chlorella, and phaeocystis, with chaetoceros having the weakest absorption. The intensity of optical absorption is proportional to the content of chlorophyll. The color of microalgae is determined by the light absorption characteristic of contained pigments (chlorophyll, phytoflavin, xanthophyll et al), which can be reflected in the transmission spectra. Therefore, the transmission spectra can reflect the color characteristic of microalgae. The transmission bands of chlamydomonas and chlorella are concentrated between 500 nm and 550 nm, indicting a green color characteristic. The transmission bands of phaeocystis and chaetoceros are concentrated between 550 nm and 625 nm, indicating a yellow color characteristic. The spectral information corresponds well with the color characteristics of the four species of microalgae and therefore, we can identify and classify the different species of microalgae through their spectrum characteristics.
Fig. 7. (a) Normalized transmission intensity spectra of the four species of microalgae and the LED source. (b) Normalized transmissivity spectra. (c) Absorption spectra captured by a spectrophotometer. and (d) normalized fluorescence spectra of the four species of microalgae.
To evaluate and verify the accuracy of the optical absorption characteristics measured by the CHMI in THI mode, reference measurements of the optical absorptivity of the microalgae were made using a spectrophotometer (UV 2550, SHIMADZU, Japan). Microalgae solutions with identical concentrations were placed in reaction cups in the spectrophotometer for characterization. The wavelength range for detection was set to 400–800 nm, which is the same as the operation wavelength range of the CHMI. The results are shown in Fig. 7(c), where two absorption peaks around 450 nm and 670 nm are seen for all four species of microalgae, i.e., the measured results of our system are consistent with the spectrophotometer characterization results, which demonstrates the accuracy of the THI mode for transmission spectrum measurements.
The main fluorescence peaks seen in Fig. 7(d) are all centered around 670 nm with additional peaks located at 725 nm, and these are consistent with the fluorescence characteristics of chlorophyll. However, due to the difference of chlorophyll content of the different species, there are differences between their fluorescence spectra. Therefore, the fluorescence spectrum information can also be used as a basis for microalgae species classification.
3.2 Classification of microalgae species based on a SVM method
In this section, we discuss the use of a statistical approach based on support vector machines (SVM) to classify different species of microalgae and compare the classification performance for various kinds of datasets. Using the CHMI we collected 300 sample sets of transmission spectra data and 300 sample sets of fluorescence spectra data for each species of microalgae. In total, for the four species of microalgae, 1200 sample sets of transmission spectra data and 1200 sample sets of fluorescence spectra data were obtained, which formed the datasets in the classification experiments.
Principal component analysis (PCA) is a common analysis method that can effectively reduce data redundancy and maximize data variance, and it is widely used in the classification of biological samples [24]. The size of the raw data matrix of the collected sample sets was 1200×1216, where the row number 1200 represents the total number of microalgae samples (300×4), and the column number 1216 represents the number of spectral sampling points for each kind of microalgae. After dimensionality reduction using the PCA approach, the original hyperspectral data is transformed into a new set of variables, termed principal components (PCs). The first few PCs typically accounts for the majority of the data variance. Here we used the first three PCs to rebuild a reduced dimensional matrix, with the size of 1200×3. The transmission and fluorescence hyperspectral spectrum datasets of the four species of microalgae were projected on a three-dimensional space constructed from the first, second and third principal components (PC1-PC3), shown in Fig. 8. The percentage in brackets indicates the proportion of the principal component. Figure 8(a) shows that the transmission hyperspectral data have a rather random distribution in the principal component space, which means that it is hardly possible to distinguish the different microalgae species using PCA of the transmission hyperspectral data. Figure 8(b) shows the distribution of fluorescence hyperspectral data in the principal component space and we can see that the data of the same microalgae species cluster together. Fluorescence hyperspectral data of different microalgae species can be segmented well in the principal component space.
Fig. 8. Distribution of (a) transmission hyperspectral data and (b) fluorescence hyperspectral data of the four species of microalgae in the principal component space.
After the feature extraction of the original data by PCA, we used SVMs to classify the data according to their labels. SVM is a supervised technique that can be used to select decision boundaries and discriminate between groups [25–27]. Here, we used a linear SVM algorithm to classify the microalgae species.
Based on the PCA-SVM analysis approach, we studied the effectiveness of the data types obtained in different imaging modes of the CHMI system for microalgae classification accuracy. We used three categories of datasets for classification: transmission hyperspectral data only, fluorescence hyperspectral data only, and combined transmission hyperspectral and fluorescence hyperspectral data. We performed PCA dimensional reduction to transform these to PC variables in the same manner as above, and the first three PCs were used as input datasets in the classification. Then, a linear SVM algorithm was used to distinguish the hyperspectral data. To preliminarily display the SVM segmentation performance, the segment regions based on PC1 and PC2 of the transmission hyperspectral data only are shown in Fig. 9(a). We can see that some data points fall into the incorrect segment regions and the classification error is high at the junction of the four regions.
Fig. 9. Segmentation regions of the (a) transmission hyperspectral data only based on PC1 and PC2; (d) fluorescence hyperspectral data only based on PC1 and PC2. Confusion matrix of using (b) transmission hyperspectral data; (e) fluorescence hyperspectral data; and (g) combined transmission hyperspectral data and fluorescence hyperspectral data for 1200 samples. ROC curves of using (c) transmission hyperspectral data; (f) fluorescence hyperspectral data; (h) combined transmission hyperspectral data and fluorescence hyperspectral data.
To evaluate the classification performance, we introduce some standard metrics: accuracy, sensitivity, and specificity. Predictive accuracy is defined as the sum of true positives (TP) and true negatives (TN) divided by the sum of all examples, as expressed in Eq. (2), where FN and FP are the number of false negatives and false positives, respectively. Sensitivity is a measure of the proportion of positives correctly identified, while specificity is that of negatives correctly identified as expressed in Eqs. (3) and 4, respectively. All the three metrics can be quantified and expressed in a confusion matrix.
The confusion matrix of all transmission spectra for 1200 samples is shown in Fig. 9(b). The average accuracy, sensitivity, and specificity to distinguish one species from the others are 74%, 74% and 89.66%, respectively, i.e., not sufficient for accurate classification. The receiver operating characteristic (ROC) curves are plotted as a TP rate versus FP rate (or 1 - specificity) in Fig. 9(c), and the area under curve (AUC) for each classification scenario are 0.94, 0.93, 0.92 and 0.97, respectively.
Figure 9(d) shows the linear segmentation regions based on PC1 and PC2 of the fluorescence hyperspectral only data. The data points of each microalga species are overall well clustered in distinct segmentation regions, however, some data points of phaeocystis erroneously fall in the regions of chlamydomonas. As seen in Fig. 7(d), the fluorescence spectra of the two above species are quite similar, which explains the large classification errors. The confusion matrix is shown in Fig. 9(e) and the average accuracy, sensitivity, and specificity are 90.08%, 90.08% and 96.67%, respectively. The ROC curves, with AUCs of 0.93, 0.91, 1.00 and 1.00, are plotted in Fig. 9(f).
To improve the classification accuracy, we combine the transmission with fluorescence hyperspectral data as a single PCA-SVM training dataset. As the dimension of the dataset is high, it is challenging to display the SVM segmentation regions on a two-dimensional plane and this is thus omitted in Fig. 9. Figure 9(g) shows the confusion matrix and the average accuracy, sensitivity and specificity were calculated to be 96.25%, 96.25% and 98.75%, respectively. The ROC curves are shown in Fig. 9(h), with AUCs of 0.99, 0.99, 1.00 and 1.00, respectively.
The classification accuracy of the transmission hyperspectral data is only 74%, and that of fluorescence hyperspectral data is 90.08%, neither of which is high enough. In the segmentation picture of transmission hyperspectral data shown in Fig. 9(a), we see that phaeocystis and chlamydomonas are well distinguished, while chaetoceros and chlorella are not. However, as seen in Fig. 9(d), chaetoceros and chlorella can be well classified, while phaeocystis and chlamydomonas were poorly so. Therefore, for classification of microalgae species, transmission and fluorescence hyperspectral data can play complementary roles as they reflect two independent types of microalgae characteristics. By combining transmission and fluorescence hyperspectral data in the PCA-SVM approach, the classification accuracy was greatly improved to 96.25%. The results thus imply that the CHMI proposed in this paper, which can collect both transmission and fluorescence hyperspectral data using two imaging modes, should prove a useful tool for biological classification.
We have demonstrated the feasibility for microalgae classification based on combined transmission and fluorescence hyperspectral data. Additionally, species identification of a group of mixed microalgae using the above method was further carried out. Four different microalgae solutions in equal amounts were placed in a same container, forming a group of mixed microalgae. The mixed solution was shaken and dropped onto a glass slide for detection. Figure 10(a) shows the transmission hyperspectral image of the mixed microalgae sample, in which the specific species can be hardly distinguished. Threshold segmentation and edge detection were carried out to identify microalgae from image background, and the processed image is shown in Fig. 10(b). Then, the transmission and fluorescence hyperspectral data of all microalgae were extracted and stored in the matrix as the test data. Using the PCA-SVM method, the identification results can be obtained, as shown in Fig. 10(c). According to the classification results, the four species of microalgae, chlamydomonas, phaeocystis, chlorella, chaetoceros, were rendered in blue, red, green and purple color, respectively.
Fig. 10. (a) Transmission hyperspectral image of the mixed microalgae. (b) The image after threshold segmentation and edge detection. (c) The image with a color mask according to the classification results by the PCA-SVM method.
3.3 Classification of microalgae species based on deep learning
Although the classification accuracy of microalgae species using PCA-SVM reached 96.3%, there is still potential for further improvement. Machine learning and deep learning are often used for the traditional inverse/analysis problems [30] and hyperspectral images classification [31–34]. They are applied to study the intrinsic relationship between samples, and extracting information about the differences among samples for classification. In this section, we used deep learning for classification analysis of different species of microalgae. Figure 11 shows the framework of a deep learning neural network (DNN). The number of hidden layers is 8 and the number of neurons in the hidden layers are [32 16 32 16 8 4 8 4]. To prevent the training network model from overfitting and to validate the performance, we randomly divided the datasets into a training set, a validation set, and a test set with ratios 0.70/0.15/0.15. The learning rate is 10−4, and the activation function is a sigmoid function. The loss function is defined as the mean square error (MSE) between the network output and the sample label. The training goal of the DNN is to minimize the MSE.
To study the influence of data types on the accuracy of microalgae classification based on DNN, we utilized three types of datasets to train and test the DNN. In the same manner as the PCA-SVM classification above, the datasets were categorized into three types: transmission hyperspectral data only, fluorescence hyperspectral data only, and combined transmission hyperspectral and fluorescence hyperspectral data. After dimensionality reduction using the PCA approach, the data sets can be used as experimental data for DNN and Fig. 12 shows the training process of the DNN based on these. With the increase of epoch, the network parameters are gradually optimized and the MSE gradually decreases. The MSE of the DNN trained by the combined transmission and fluorescence hyperspectral data reached the minimum MSE of the three data sets.
Fig. 12. DNN training process based on (a) transmission hyperspectral data, (b) fluorescence hyperspectral data, (c) combined transmission hyperspectral data and fluorescence hyperspectral data.
Figs. 13(a-c) show the consistency between the output of the DNN trained by transmission hyperspectral data only and sample labels. From left to right are the results based on the training set, validation set and test set. The green area indicates the correct part of the classification and the classification accuracies of the three sets are 83.19%, 82.32%, and 81.42%, respectively. Figs. 13(d-f), shows the same for the fluorescence hyperspectral data only, for which the classification accuracy of its training set, validation set, and test set are 96.83%, 95.74%, and 95.52%, respectively. Figs. 13(g-j) then show the consistency between the output of the DNN trained by the combined data and sample labels. The classification accuracies of its training set, validation set, and test set are 99.67%, 98.76%, and 98.34%, respectively. The results again show the advantage of combining transmission and fluorescence hyperspectral data as it can greatly improve the performance of classification of microalgae species. Compared with the PCA-SVM method, the DNN further improves the classification accuracy from 96.3% to 98.34%, implying that this approach could be useful for a wider range of biological classification applications.
Fig. 13. Consistency between the output of the DNN and sample labels. Input datasets are: (a) training, (b) validation and, (c) test sets using transmission hyperspectral data only; (d) training, (e)validation, and (f) test sets using fluorescence hyperspectral data only; (g) training, (h) validation and (j) test sets using combined fluorescence and transmission hyperspectral data.
3.4 Hyperspectral imaging experiments of jellyfish with symbiotic microalgae
Microalgae are not only distributed in the ocean in a free state, but also form symbionts with many marine organisms, such as jellyfish, corals, sea anemones, and the green hydra. Symbiotic microalgae and symbionts coexist mutualistically. Symbiotic microalgae are mainly autotrophs, which can provide the symbiont with translocated reducing carbides, such as products of photosynthesis glucose, glycerol, amino acids, etc. In other words, symbiotic microalgae play a key role in the lifecycle and nutrient uptake of symbionts and thus observation of symbiotic microalgae is of great significance. However, due to the micro size and weak optical absorption of symbiotic microalgae, traditional hyperspectral microscope systems are not suitable for extraction of useful spectral information. The low resolution of traditional systems furthermore makes it more difficult to identify locations and distribution of the microalgae.
In contrast to conventional optical microscope systems, the CHMI, which can acquire both fluorescence and transmission hyperspectral images, is well suited for analysis of the optical properties of symbiotic microalgae in symbionts as well as to map microalgae distributions. The transmission or fluorescence spectra can provide much inherent optical information of the samples and the fluorescence images can in addition clearly show the distribution of microalgae. To demonstrate this, we use the CHMI to examine four different species of jellyfish: brown cassiopea andromeda jellyfish, blue cassiopea andromeda jellyfish, and edible jellyfish.
First, we examined a live specimen of brown cassiopea andromeda jellyfish, for which we used a 5× microscope objective (OLYMPUS Inc., LUMPLFLW series) in the CHMI. Figure 14(a) shows a microscopic image of the jellyfish, in which the oral arm (blue rectangle), gastric pouch (green rectangle), and mesoglea (red rectangle) are indicated. The transmission hyperspectral images of the indicated regions are shown in Figs. 14(b-d), where the black areas are regions that has strong optical absorption caused by microalgae or physiological materials of the jellyfish. The corresponding fluorescence hyperspectral images are shown in Figs. 14(e-g) in which the distributions of the microalgae are clear from the excitation of fluorescent light by the chlorophyll. In Fig. 14(e), we can see that the microalgae cluster at the edges of the oral arm as these areas exhibit strong excited fluorescent light. From Fig. 14(g) we can deduce a sparse distribution of microalgae at the mesoglea as the intensity of fluorescent light is weaker than that in Figs. 14(e) and (f). The normalized transmission intensity spectra from points A to F in the oral arm edge, gastric pouch, and mesoglea indicated in Fig. 14(b-d) are shown in Figs. 15(a-c), and the normalized transmissivity spectra (intensity spectrum of the jellyfish divided by that of the LED) in Figs. 15(d-f). Points B, D, and E are the microalgae located at oral arm edge, gastric pouch, and mesoglea, respectively, while points A, C, and F are the position of the corresponding parts showed no fluorescence characteristics. The transmissivity spectrum curves of points B, D and E have two absorption peaks at 450 nm and 670 nm that corresponds to the optical characteristics of chlorophyll. The transmissivity at 450 nm and 670 nm at point E is lower than that at points B and D, indicating that the microalgae content is lower at the mesoglea, which is consistent with the conclusion from the fluorescence images in Fig. 14. As seen in Figs. 15(e) and (f), transmissivity of points C and F in the 660–700 nm range is much higher than in the 400–600 nm range with an upward trend, which reflects the brown color characteristics of the jellyfish. Figs. 15(g-j) show the normalized fluorescence spectra of points B, D, and E, respectively. The main fluorescence spectrum peaks are located at 680 nm with an additional peak at 720 nm. Both the normalized transmissivity spectra and the normalized fluorescence spectra of the three points are similar in shape, indicating that the microalgae found at the three locations of are of the same species.
Fig. 14. (a) Microscopic image of brown cassiopea andromeda jellyfish. Transmission hyperspectral images of the (b) blue (c) green, and (d) red rectangular regions indicated in (a); Fluorescence hyperspectral images of the (e) blue, (f) green, and (g) red rectangular regions indicated in (a).
Fig. 15. Normalized transmission intensity spectra of (a) points A and B in Fig. 14(b); (b) points C and D in Fig. 14(c); (c) points E and F in Fig. 14(d). Normalized transmissivity spectra of (d) points A and B; (e) points C and D; (f) points E and F. Normalized fluorescence spectrum of (g) point B; (h) point D; and (j) point E.
Imaging experiments of blue cassiopea andromeda jellyfish were performed in the same manner as above and the results are displayed in the same fashion in Figs. 16 and 17. Compared with the brown cassiopea andromeda jellyfish, the fluorescence intensities of the blue cassiopea andromeda jellyfish are much weaker, and the fluorescence regions are smaller, indicating that less microalgae are attached to the blue cassiopea andromeda jellyfish. Additionally, as seen in Fig. 16(g), microalgae distributed on the mesoglea of jellyfish are mainly in free form, shown as single point fluorescence distribution. Points A, C and G indicated in Figs. 16(b-d) are microalgae located at the oral arm edge, the gastric pouch, and the mesoglea, respectively, while points B, D, E and F are the position of the corresponding parts showed no fluorescence characteristics. As seen in Figs. 17(d) and (e), the transmissivity spectra curves of points A and C have two absorption peaks at 450 nm and 670 nm, i.e., corresponding to characteristic of chlorophyll. In contrast to the corresponding spectra for the brown cassiopea andromeda jellyfish, its optical absorption at 450 nm is significantly larger, while it is smaller at 670 nm, indicating that the species of the symbiotic microalgae of the two jellyfish are different. The transmissivity spectrum of point G does not reflect chlorophyll characteristic, which is due to the fact that the microalgae at point G is in a free form and optical absorption is too weak to be detected by the hyperspectral imager. The transmissivity spectra of points B and D in Fig. 17(d) and (e) reflect no microalgae distribution, which is confirmed by the fluorescence images in Fig. 16(e) and (f) where points B and D are black as well. Points E and F are located at the radial canals and enteric cavity of the mesoglea. According to the spectra curves in Fig. 17(f), there are no microalgae at these two parts. The transmissivity spectrum of point E has a transmission peak between 450 nm and 550 nm, reflecting the blue color characteristic of the jellyfish. Figs. 17(g-j) show the normalized fluorescence spectra of points A, C, and G, respectively. These are similar in shape, corresponding to the fluorescence characteristic of chlorophyll. While the chlorophyll of microalgae at point G could not be detected using the transmission spectrum as discussed above, it clearly can be detected and characterized by the fluorescence spectrum, a further demonstration of the advantage of combined fluorescence and transmission spectrum characterization.
Fig. 16. (a) Microscopic image of blue cassiopea andromeda jellyfish. Transmission hyperspectral images of the (b) blue (c) green, and (d) red rectangular regions indicated in (a); Fluorescence hyperspectral images of the (e) blue, (f) green, and (g) red rectangular regions.
Fig. 17. Normalized transmission intensity spectra of (a) points A and B in Fig. 16(b); (b) points C and D in Fig. 16(c); (c) points E, F and G in Fig. 16(d). Normalized transmissivity spectra of (d) points A and B; (e) points C and D; (f) points E, F and G. Normalized fluorescence spectrum of (g) point A; (h) point C; (j) point G.
As a final demonstration measurement, we performed hyperspectral imaging on a specimen of edible jellyfish. A Microscopic image is shown in Fig. 18(a) and transmission and fluorescence hyperspectral images of the region indicated by a red rectangle are shown in Figs. 18(b) and (c). The fluorescence hyperspectral image does not show any excited fluorescence, indicating that there are no symbiotic microalgae. Points A and B located at the body and oral arm of the specimen, respectively. The normalized transmission intensity, normalized transmissivity of points A and B indicated Fig. 18(b) are shown in Figs. 18(d-e). The transmissivity spectrum of point B has a transmissivity peak between 580 nm and 610 nm, which corresponds to the red color of the jellyfish’s tentacles. The transmissivity spectrum of point A and its spectrum collected in CFHI mode (as seen in Fig. 18(f)) do not have optical absorption or fluorescence characteristic of chlorophyll, thus supporting the conclusion that the edible jellyfish does not contain symbiotic microalgae.
Fig. 18. (a) Microscopic image of edible jellyfish. (b) Transmission and (c) fluorescence hyperspectral images of the region indicated by the red rectangle in (a). (d) Normalized transmission intensity spectrum of points A, B and the LED source. (e) Transmissivity spectrum of points A and B. (f) Fluorescence spectrum of point B.
4. Summary and outlook
In this paper, we propose a confocal hyperspectral microscopic imager (CHMI) that integrates a confocal fluorescence hyperspectral imaging (CFHI) mode and a transmission hyperspectral imaging (THI) mode. In the CFHI mode, a confocal technique is used to eliminate image blurring caused by the interference of axial fluorescent spots in a conventional fluorescence hyperspectral imager. Consequently, the fluorescence spectrum of single fluorescent point and a high-resolution fluorescence hyperspectral image can be obtained. The CHMI works in staring state, and the two imaging modes share the same main optical path and therefore there is a point-to-point correspondence between the fluorescence and transmission images. The combination of two types of spectral information yields superior performance in characterizing and distinguishing the optical properties differences of micro-sized samples. The CHMI shows excellent performance, with spectral and spatial resolutions of 3 nm and 2 µm, respectively.
We have presented a set of experiments on microalgae in free state in which the CHMI collected the transmission and fluorescence spectra of four different species of microalgae for classification. After PCA pre-processing, SVM and deep learning were used for classification, for which classification accuracies of 96.3% and 99.67%, respectively, were reached. The results indicate that the combined THI and CFHI modes can realize high classification accuracy of different microalgae species in low concentrations. We furthermore demonstrated the ability of the CHMI to analyze concentration, species, and distribution differences of symbiotic microalgae in symbionts. The CHMI thus shows significant potential for applications in a broad range of areas, such as microbiological imaging, health and growth state assessment of symbionts, and on-site marine environmental monitoring.
Areas for future system improvement includes a wider wavelength detection range, hyperspectral imaging combined with three-dimensional (3D) reconstruction, and polarized hyperspectral analysis, all of which are feasible and under consideration. The CHMI can be extended to cover a wider spectral range including ultraviolet (UV) and infrared (IR), providing rich fingerprint features [35–37]. For example, the infrared spectrum can be used to measure the water content of samples [38]. Hyperspectral imaging combined with 3D reconstruction [15] would provide more comprehensive spatial information of the samples and by combining hyperspectral detection technology with polarization analysis, object identification measurement could be more precise.
Funding
National Key Research and Development Program of China (2018YFC1407503); Key Research and Development Program of Zhejiang Province (2021C03178); Ningbo Science and Technology Project (2020Z077, 20211ZDYF020103); Ningbo Science and Technology Plan Project-Key Core Technology Emergency Tackling Plan Project (2020G012).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
References
1. J. P. Fidalgo, A. Cid, E. Torres, A. Sukenik, and C. Herrero, “Effects of nitrogen source and growth phase on proximate biochemical composition, lipid classes and fatty acid profile of the marine microalga Isochrysis galbana,” Aquaculture 166(1-2), 105–116 (1998). [CrossRef]
2. Z. Xu, Y. Jiang, J. Ji, E. Forsberg, Y. Li, and S. He, “Classification, identification, and growth stage estimation of microalgae based on transmission hyperspectral microscopic imaging and machine learning,” Opt. Express 28(21), 30686–30700 (2020). [CrossRef]
3. C. Bumbac, E. Manea, A. Banciu, C. Stoica, I. Ionescu, V. Badescu, and M. N. Lazar, “Identification of Physical, Morphological and Chemical Particularities of Mixed Microalgae - Bacteria Granules,” Rev. Chim. 70(1), 275–277 (2019). [CrossRef]
4. A. M. Marba-Ardebol, J. Emmerich, M. Muthig, P. Neubauer, and S. Junne, “In Situ Microscopy for Real-time Determination of Single-cell Morphology in Bioprocesses,” J. Vis. Exp. 2019(154), 1–9 (2019). [CrossRef]
5. J. M. Bioucas-Dias, A. Plaza, G. Camps-Valls, P. Scheunders, N. M. Nasrabadi, and J. Chanussot, “Hyperspectral Remote Sensing Data Analysis and Future Challenges,” IEEE Geosci. Remote Sens. Mag. 1(2), 6–36 (2013). [CrossRef]
6. N. Caporaso, M. B. Whitworth, S. Grebby, and I. D. Fisk, “Non-destructive analysis of sucrose, caffeine and trigonelline on single green coffee beans by hyperspectral imaging,” Food Res. Int. 106, 193–203 (2018). [CrossRef]
7. X. Yao, S. Li, and S. He, “Dual-Mode Hyperspectral Bio-Imager with a Conjugated Camera for Quick Object-Selection and Focusing,” Progress in Electromagnetics Research-Pier 168, 133–143 (2020). [CrossRef]
8. M. Zhu, D. Huang, X.-J. Hu, W.-H. Tong, B.-L. Han, J.-P. Tian, and H.-B. Luo, “Application of hyperspectral technology in detection of agricultural products and food: A Review,” Food Science & Nutrition 8(10), 5206–5214 (2020). [CrossRef]
9. Y. Liu, H. Pu, and D.-W. Sun, “Hyperspectral imaging technique for evaluating food quality and safety during various processes: A review of recent applications,” Trends Food Sci. Technol. 69, 25–35 (2017). [CrossRef]
10. G. Notesco, S. Weksler, and E. Ben-Dor, “Mineral Classification of Soils Using Hyperspectral Longwave Infrared (LWIR) Ground-Based Data,” Remote Sens. 11(12), 1429 (2019). [CrossRef]
11. L. Coic, P.-Y. Sacre, A. Dispas, A. K. Sakira, M. Fillet, R. D. Marini, P. Hubert, and E. Ziemons, “Comparison of hyperspectral imaging techniques for the elucidation of falsified medicines composition,” Talanta 198, 457–463 (2019). [CrossRef]
12. G. Lu and B. Fei, “Medical hyperspectral imaging: a review,” J. Biomed. Opt. 19(1), 010901 (2014). [CrossRef]
13. S. Zhu, K. Su, Y. Liu, H. Yin, Z. Li, F. Huang, Z. Chen, W. Chen, G. Zhang, and Y. Chen, “Identification of cancerous gastric cells based on common features extracted from hyperspectral microscopic images,” Biomed. Opt. Express 6(4), 1135–1145 (2015). [CrossRef]
14. J. Luo, F. Cai, X. Yao, J. Li, Q. Huang, and S. He, “Experimental Demonstration of an Anti-Shake Hyperspectral Imager of High Spatial Resolution and Low Cost,” IEEE Sens. J. 20(14), 8082–8090 (2020). [CrossRef]
15. J. Luo, S. Li, E. Forsberg, and S. He, “4D surface shape measurement system with high spectral resolution and great depth accuracy,” Opt. Express 29(9), 13048–13070 (2021). [CrossRef]
16. H. Lin, Y. Zhang, and L. Mei, “Fluorescence Scheimpflug LiDAR developed for the three-dimension profiling of plants,” Opt. Express 28(7), 9269–9279 (2020). [CrossRef]
17. M. A. Ilie, C. Caruntu, D. Lixandru, M. Tampa, S.-R. Georgescu, M.-M. Constantin, C. Constantin, M. Neagu, S. A. Zurac, and D. Boda, “In vivo confocal laser scanning microscopy imaging of skin inflammation: Clinical applications and research directions,” Exp. Ther. Med. 17(2), 1004–1011 (2019). [CrossRef]
18. M. A. Ilie, C. Caruntu, M. Lupu, D. Lixandru, M. Tampa, S.-R. Georgescu, A. Bastian, C. Constantin, M. Neagu, S. A. Zurac, and D. Boda, “Current and future applications of confocal laser scanning microscopy imaging in skin oncology,” Oncol. Lett. 17(5), 4102–4111 (2019). [CrossRef]
19. J. Li, F. Cai, Y. Dong, Z. Zhu, X. Sun, H. Zhang, and S. He, “A portable confocal hyperspectral microscope without any scan or tube lens and its application in fluorescence and Raman spectral imaging,” Opt. Commun. 392, 1–6 (2017). [CrossRef]
20. R. H. Webb, “Confocal optical microscopy,” Rep. Prog. Phys. 59(3), 427–471 (1996). [CrossRef]
21. M. Taniguchi and J. S. Lindsey, “Absorption and Fluorescence Spectral Database of Chlorophylls and Analogues,” Photochem. Photobiol. 97(1), 136–165 (2021). [CrossRef]
22. S. Nozue, M. Katayama, M. Terazima, and S. Kumazaki, “Comparative study of thylakoid membranes in terminal heterocysts and vegetative cells from two cyanobacteria, Rivularia M-261 and Anabaena variabilis, by fluorescence and absorption spectral microscopy,” Biochimica Et Biophysica Acta-Bioenergetics 1858(9), 742–749 (2017). [CrossRef]
23. E. Baszanowska and Z. Otremba, “Spectral signatures of fluorescence and light absorption to identify crude oils found in the marine environment,” Journal of the European Optical Society-Rapid Publications 9, 14029 (2014). [CrossRef]
24. S. Wold, K. Esbensen, and P. Geladi, “PRINCIPAL COMPONENT ANALYSIS,” Chemom. Intell. Lab. Syst. 2(1-3), 37–52 (1987). [CrossRef]
25. F. Melgani and L. Bruzzone, “Classification of hyperspectral remote sensing images with support vector machines,” IEEE Trans. Geosci. Remote Sens. 42(8), 1778–1790 (2004). [CrossRef]
26. D. K. Pathak, S. K. Kalita, and D. K. Bhattacharya, “Hyperspectral image classification using support vector machine: a spectral spatial feature based approach,” Evol. Intel. (2021). [CrossRef]
27. W. Yang and H. Song, “Spectral-Spatial Classification of Hyperspectral Image Based on Support Vector Machine,” Int. J. Inf. Technol. Web. Eng. 16(1), 56–74 (2021). [CrossRef]
28. J. H. Cho, P. J. Gemperline, and D. Walker, “Wavelength calibration method for a CCD detector and multichannel fiber-optic probes,” Appl. Spectrosc. 49(12), 1841–1845 (1995). [CrossRef]
29. A. K. Katsaggelos, J. Biemond, R. M. Mersereau, and R. W. Schafer, “AN ITERATIVE METHOD FOR RESTORING NOISY BLURRED IMAGES,” Circuits Systems and Signal Processing 3(2), 139–160 (1984). [CrossRef]
30. X. Chen, Z. Wei, M. Li, and P. Rocca, “A Review of Deep Learning Approaches for Inverse Scattering Problems,” Progress in Electromagnetics Research-Pier 167, 67–81 (2020). [CrossRef]
31. W. Zhao and S. Du, “Spectral-Spatial Feature Extraction for Hyperspectral Image Classification: A Dimension Reduction and Deep Learning Approach,” IEEE Trans. Geosci. Remote Sens. 54(8), 4544–4554 (2016). [CrossRef]
32. Y. Chen, Z. Lin, X. Zhao, G. Wang, and Y. Gu, “Deep Learning-Based Classification of Hyperspectral Data,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 7(6), 2094–2107 (2014). [CrossRef]
33. H. Lin, Z. Li, H. Lu, S. Sun, F. Chen, K. Wei, and D. Ming, “Robust Classification of Tea Based on Multi-Channel LED-Induced Fluorescence and a Convolutional Neural Network,” Sensors 19(21), 4687 (2019). [CrossRef]
34. T. Ma, H. Lyu, J. Liu, Y. Xia, C. Qian, J. Evans, W. Xu, J. Hu, S. Hu, and S. He, “Distinguishing Bipolar Depression from Major Depressive Disorder Using fNIRS and Deep Neural Network,” Progress in Electromagnetics Research-Pier 169, 73–86 (2020). [CrossRef]
35. Z. Yong, S. Zhang, Y. Dong, and S. He, “Broadband nanoantennas for plasmon enhanced fluorescence and Raman spectroscopies,” Prog. Electromagn. Res. 153, 123–131 (2015). [CrossRef]
36. G. Wang, P. Kulinski, P. Hubert, A. Deguine, D. Petitprez, S. Crumeyrolle, E. Fertein, K. Deboudt, P. Flament, M. W. Sigrist, H. Yi, and W. Chen, “Filter-free light absorption measurement of volcanic ashes and ambient particulate matter using multi-wavelength photoacoustic spectroscopy,” Prog. Electromagn. Res. 166, 59–74 (2019). [CrossRef]
37. D. DePaoli, É Lemoine, K. Ember, M. Parent, M. Prud’homme, L. Cantin, K. Petrecca, F. Leblond, and D. C. Côté, “Rise of Raman spectroscopy in neurosurgery: a review,” J. Biomed. Opt. 25(05), 1 (2020). [CrossRef]
38. M. Corti, P. Marino Gallina, D. Cavalli, and G. Cabassi, “Hyperspectral Imaging of Spinach Canopy under Combined Water and Nitrogen Stress to Estimate Biomass, Water, and Nitrogen Content,” Biosyst. Eng. 158, 38–50 (2017). [CrossRef]
