Open Access
Add to BibSonomy
Abstract
Surface-enhanced Raman scattering (SERS) spectroscopy analysis has long been the central task of nanoscience and nanotechnology to realize the ultrasensitive recognition/quantitation applications. Recently, the blooming of artificial intelligence algorithms provides an edge tool to revolutionize the spectroscopy analysis, especially for multiple substances analysis and large-scale data handling. In this study, a single-model multi-tasks deep learning network is proposed to simultaneously achieve the qualitative recognition and quantitative analysis of SERS spectroscopy. The SERS spectra of two kinds of hypoglycemic drugs (PHE, ROS) and the corresponding mixtures are collected, respectively, with the concentration grade from 10−4 M to 10−8 M. Based on the SERS spectroscopy dataset, the loss functions and hyperparameters of the multi-tasks classifications model are optimized, and the recognition accuracies are tested by simulation experiments. It is demonstrated that the accuracy rates of qualitative and quantitative analysis can reach up to 99.0% and 98.4%, respectively. Moreover, the practical feasibility of this multi-tasks model is demonstrated by using it to achieve qualitative and quantitative analysis of PHE and ROS in complex serum matrix. Overall, this single-model multi-tasks deep learning network shows significant potential for the recognition and quantitation of SERS spectroscopy, which provides the algorithmic and experimental basis for large-scale and multiple substances SERS spectra analysis.
© 2022 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Surface-enhanced Raman scattering (SERS) spectroscopy has been intensely explored as a powerful analytical technique with applications in chemical science, bio-analysis and environmental/food safety [1–4]. On one hand, benefiting from the plasmonic enhancement, SERS can provide abundant vibrational fingerprints information about the molecular structure and composition. Especially, when molecules located in the gap of metallic nanostructures (“hot spot”), the enormous near-field enhancement can make single-molecule SERS detection possible [5–8]. With the advantages of molecular specificity and sensitivity, SERS has realized ultrasensitive detection for applications related to drugs examination, pesticides residue analysis and environmental monitoring. On the other hand, quantitative SERS analysis has long been a central task of nanoscience and nanotechnology. By improving the fabrication of substrates or introducing internal/external standard, the SERS quantitative technology has recently made a great progress [9–14]. For instance, three-dimensional metal nanostructures of uniform size have been designed to accomplish a good quantitative detection [15–17]. And by introducing the internal intensity standard, quantitative SERS detection down to 3 × 10−11 M has been achieved [18–20]. Based on this concept, SERS has the potential to be a powerful nondestructive analytical technique for routine qualitative recognition (type of the sample) and quantitative analysis (concentration of the sample). Traditionally, for SERS spectroscopy analysis, the substance is first recognized through the attribution and comparison of characteristic peaks. Then, the peaks intensity is extracted from the SERS spectra to fit the “intensity-concentration” calibration curve for quantitative analysis. However, these processes are always tedious and require experienced personnel, especially for multiple substances detection, which limit the effectiveness and populations of SERS applications.
Recently, the blooming of artificial intelligence and data mining technology provides an edge tool for data analysis. As one of the main branches of machine learning, deep learning takes the advantages of pre-training and fine-tuning mechanisms, and has been widely applied in many fields such as recognition and classification [21–24]. In recent years, more and more works resort to deep learning to seek breakthrough in spectral analysis [25–27]. These works have demonstrated good performance in their specific task scenarios. For example, Hao and Hu et al. used convolutional neural network (CNN) models to realize infrared spectra and terahertz time-domain spectra recognition [28,29]. And the accuracy rates on the test set reached 91.2% and 99.6%, respectively. However, in these works, the one-dimensional spectral data needed to be converted into two-dimensional images to realize the classification, which not only brought more complicated data preprocessing steps, but also destroyed the original structure of the spectral data. In our previous work, the improved 2D convolution ResNet [30] was used to solve the problem of one-dimensional Raman spectral data analysis [31]. In five sets of cross-validation experiments, the recognition accuracy of the test set reached 100%. While, this work has only completed the qualitative analysis of spectral species recognition, the quantitative analysis has not been accomplished. Based on the above problems of spectral analysis, we design a single-model multi-tasks architecture. The improved ResNet is used as the backbone to extract spectral features, and two classifiers are trained for spectral recognition and quantitative analysis at the same time, so that we can use only one model to complete the recognition and quantitative analysis of the spectra simultaneously.
In this paper, phenformin hydrochloride (PHE) and rosiglitazone maleate (ROS) were chosen as analytes to demonstrate the effectiveness of deep learning model in SERS spectroscopy analysis. These two hypoglycemic drugs have always been illegally added in health foods to lower blood sugar. Schemes of molecular structures are shown in Supplement 1 (Scheme S1). The abuse of them might lead to serious side effects such as anemia, cardiovascular injury, lactic acidosis and liver/kidney damage. As stipulated by the Food and Drug Administration (FDA), PHE and ROS were banned in health foods, and most public hospitals have also eliminated these drugs [32,33]. Hence, a facile qualitative recognition and quantitative analysis of them is necessary. Firstly, a homogeneous, stable and reproducible SERS substrate was prepared by Langmuir–Blodgett method. And the SERS spectra of PHE, ROS and mixtures of them were measured, respectively, at the concentration grades from 10−4 M to 10−8 M. Then, we built a single-model multi-tasks deep learning architecture to complete the qualitative and quantitative analysis of SERS spectroscopy. Experiments show that our method has achieved good results in both qualitative and quantitative analysis with the accuracy up to 99.0% and 98.4%, respectively. It is anticipated that this single-model multi-tasks deep learning model can be easily transplanted into portable Raman spectrometers in the future, which provides the algorithmic and experimental basis for large-scale, on-site SERS spectra analysis.
2. Materials and methods
2.1 Materials
Silver nitrate (AgNO3, 99.99%) and sodium citrate (99%) were purchased from Sigma-Aldrich. The phenformin hydrochloride (PHE, 97%), rosiglitazone maleate (ROS, ≥ 99%), acetone (99.5%), toluene (99.5%), ethanol (99.5%) and methanol (99.5%) were all purchased from Innochem Technology Co., Ltd. (Beijing, China). All of these reagents were used without further purification. Milli-Q ultrapure water was used throughout this work, unless otherwise stated.
2.2 Apparatus
The scanning electron microscope (SEM) images were obtained using a field-emission scanning electron microscope (Hitachi S4800). SERS measurements were performed on an Invia Renishaw Raman spectrometer at the excitation of a 633 nm laser. The excitation power was 0.1 mW, and the integration time was 10 s.
2.3 Preparation of the SERS substrate
Synthesis of Ag NPs. The silver nanoparticles (Ag NPs) colloid was synthesized by referring to Lee-Meisel method [34]. Typically, 90 mg of AgNO3 was dissolved in 500 mL boiling water, and 10 mL of 1% sodium citrate solution was added dropwise under vigorous stirring over a period of ∼45 minutes. The mixture was then kept boiling for ∼30 minutes with continued stirring and finally cooled down to room temperature, which resulted in the formation of Ag NPs with 60 nm in diameter approximately.
Monolayer Assembly of Ag NPs. The preparation of Ag NPs film is shown in Fig. 1. Firstly, the beaker was washed with acetone to remove residual contaminants. Then, a total of 5 mL toluene and 4.5 mL of ethanol were added to the Ag NPs colloid to form a toluene/water interface. Due to the effect of the interfacial tension, the Ag NPs monolayer film (MLF) was formed within 2 minutes and was kept undisturbed for 30 minutes to volatilize the toluene. Finally, the MLF was transferred onto a silicon wafer.
Fig. 1. The formation and transfer of the self-assembly Ag NPs monolayer film. (a) Rinse the beaker with acetone. (b) Ag NPs assembled at the toluene/water interface. (c) Shrinkage and transfer process of the monolayer film obtained from toluene/water interface. (d) Schematic of the monolayer film deposited on a silicon wafer.
2.4 Data collection and preparation
In this work, we prepared PHE and ROS alcoholic solutions, respectively, with five concentration grades from 10−4 M to 10−8 M. At each concentration grade, the mixture samples were prepared by mixing the PHE and ROS with concentration ratio 4:1, 1:1 and 1:4. A small drop of analyte solution (50 µL) was dispersed on the substrate to prepare SERS sample. For each sample, 100 spectra were collected randomly from different sites, and a total of 2500 spectra were acquired. Before analyzing the spectra by deep learning, the baseline calibration was done with the LabSpec software. Finally, for each type of sample, 80% spectra were collected as training set, 20% as test set. Therefore, the training set data was 2000, the test set data was 500, and the data dimension was one-dimensional spectral intensity information with a size of 708.
2.5 Model structure
The overall architecture of the single-model multi-tasks network is shown in Fig. 2. The spectral data is first input to the following two Residual modules after a 1D convolution with kernel size of 3 and a maximum pooling layer with a stride of 2. Then, after a Global Average Pooling (GAP) layer, the features extracted by the residual modules are put into a qualitative classifier (5 output nodes, referring to the types of samples) and a quantitative classifier (5 output nodes, referring to the concentrations of samples), respectively. In the Residual module shown in Fig. 2(b), the input feature X first passes through the 1D convolution layer with a convolution kernel size of 3 (Conv3). And then passes through the regularization BatchNormalization (BN) [35] and the nonlinear activation function ReLU. Repeating the above operations, the output Y1 can be expressed as:
where X1 is represented as:Fig. 2. (a) Overall architecture of single-model multi-tasks network. (b) Residual module. (c) Classifier.
Different from the backbone, in the stage of skip connection, we use a 1D convolution layer with a convolution kernel size of 1 (Conv1) to operate on the input feature X, the purpose of which is to ensure that the channel dimensions of output Y2 and Y1 remains the same
Finally, Y2 and Y1 are accumulated, and the final result Y is obtained after the operation of regularized BN and nonlinear activation function ReLU:
The features extracted from the two residual modules are subjected to GAP [36]. The main method of GAP is to average based on the channel dimension of the feature. The purpose of using it is to convert multi-dimensional features into one-dimensional features and perform some downstream tasks. Compared with the traditional reshape method of features, this method reduces the redundant information of features, and effectively improves the ability of downstream tasks of the model while reducing Macs and Parameters. Based on this, single-model multi-tasks network architecture also adopts this operation to complete the subsequent downstream tasks (qualitative recognition and quantitative analysis of SERS spectra). For the two classifiers, we both use the linear layer with the ReLU function, as shown in Fig. 2(c). Between the ReLU activation function and the second linear layer, a dropout operation is added. Dropout can make the classifier dropped some nodes during training, which prevents the trained model overfitting. Both of the two classifiers have 5 output nodes, corresponding to the types of the samples and concentration grades, respectively, as shown in Fig. 2(a).
2.6 Loss functions and other hyperparameters
Different from the typical single-task classification model, the multi-tasks classification model needs to redesign the loss function:
Table 1. Hyperparameters setting
3. Results and discussion
3.1 Characterization and performance of MLFs
SEM images of the Ag NPs MLF prepared through the Langmuir–Blodgett method are shown in Figs. 3(a), 3(b), from which it can be seen that the Ag NPs are packed closely. Next, we investigated the SERS performance of the MLF. As shown in Fig. 3(c), for PHE (10−6 M), 20 spectra were chosen randomly from the dataset, and the MLF displayed nearly uniform SERS spectra of PHE. The statistics of PHE SERS intensity at 1006 cm−1 was shown in Fig. 3(d). The small relative standard deviation (RSD) value (∼4.9%) directly proves the well SERS uniformity of the MLF substrate. This high-quality SERS substrate guarantees a high reliability of the obtained data. In addition, as shown in Figs. 3(e), 3(f), the SERS fingerprints of PHE and ROS appear to be different from each other. As for PHE (Fig. 3(e)), the main characteristic peaks at 1006 cm-1 (benzene ring breathing vibration) and 1034 cm-1 (δC-H) can be clearly observed. As for ROS, the Raman peak at 1180 cm-1 (δC-H) and 1600 cm-1 (vC = C) can be distinctly observed even down to 10−8 M [37,38]. Obviously, the SERS intensity increases with improving the concentration of both substances. In addition, the SERS spectra of mixture samples are shown in Fig. S1 in Supplement 1. The fingerprints of PHE and ROS can be simultaneously observed, but with different spectral intensity distribution depending on the concentration ratio (4:1, 1:1 and 1:4). Overall, the above results confirm the feasibility of qualitative recognition and quantitative analysis of PHE, ROS and mixtures through SERS method. However, the large-scale SERS qualitative recognition and quantitative analysis, especially for multiple substances analysis, will be difficult for manual visual inspection and consume lots of manpower. To tackle that, we try to introduce deep learning to realize rapid, accurate and large-scale SERS data recognition and quantitative analysis.
Fig. 3. (a), (b) SEM images of Ag nanoparticles MLF. (c) A series of SERS spectra of PHE from 20 different sites on the MLF substrate. (d) The SERS intensity of PHE at 1006 cm-1 for spectra in (c). (e), (f) SERS spectra of PHE and ROS at different concentrations.
3.2 Loss function weights selection
In the model, first, we adjust the weights of the two loss functions to train the optimum model for the qualitative and quantitative spectral analysis. As shown in Table 2, the values of λ1 and λ2 are from 0.1 to 0.9 and 0.9 to 0.1, respectively. The experiments demonstrate that, when λ1 = 0.5 and λ2 = 0.5, the two classifiers have the best effect on spectral qualitative recognition and quantitative analysis, reaching 99.0% and 98.4% recognition accuracies. However, with the increase of λ1, although the recognition accuracy of qualitative analysis remains good, the accuracy of quantitative analysis task gradually declines, resulting in a decrease in the average accuracy. In addition, we also tried to use regression method to deal with the quantitative analysis, where MSE was used as loss function. As shown in Table S1, the recognition accuracies of qualitative and quantitative analysis by joint training of two CE loss functions are better than that using classification CE and regression MSE loss functions (see Supplement 1 for details).
Table 2. λ1 and λ2 weights selection results
3.3 Comparisons between multi-tasks model and single-task models
As discussed above, for multi-tasks model, it is critical to select appropriate loss functions weights for different tasks to optimize the global recognition accuracies. Here, comparisons between two single-task models and multi-tasks model are also carried out. For single-task network, only one model is designed. To complete the qualitative and quantitative works, different classifiers are used, while maintains model structure and other parameters the same. The comparison results are shown in the Table 3. The recognition accuracy rate in single-task qualitative recognition is 99.2%, while the accuracy of single-task quantitative analysis is 98.6%. Both of them are slightly higher than the accuracies of multi-tasks model. Compared with single model to complete single task, our proposed network architecture for a single model to complete multiple tasks can simultaneously achieve qualitative recognition and quantitative analysis with high accuracies, which can not only reduce the training time of the required model and the amount of model parameters, but also expand the application scenarios of the deep learning model.
Table 3. The comparison results between multi-tasks model and single-task models
3.4 Qualitative recognition and quantitative analysis of PHE, ROS in serum
To demonstrate the practical feasibility of the proposed deep learning network, we further detected PHE and ROS molecules with different concentrations in Fetal Bovine Serum, respectively. Serum, consisting of amino acids, sugars, and larger neuropeptides or proteins molecules, is a highly complex matrix. Here, serum samples were prepared by adding methanol to serum, followed by centrifugation of the complicated sample to remove large protein interferences (see Supplement 1 for details). The representative SERS spectra of PHE and ROS in serum are shown in Figs. 4(a) and 4(b). In terms of the serum itself, the SERS peaks of amino acid, purine and phospholipids can be observed, such as 661 cm-1 ascribed to v(C-S) of tyrosine, 728 cm-1 from the v(C-H) of hypoxanthine, 1002 cm-1 from the C-C twisting mode of Phenylalanine [39]. Moreover, when PHE/ROS was added into the serum, the fingerprints of PHE (ROS) at 1006 cm-1, 1034 cm-1 (1180 cm-1) and serum peaks can be observed simultaneously. And the intensity of PHE/ROS decreases with the reducing of concentration from 10−4 M to 10−6 M. For each sample, 100 SERS spectra were acquired randomly from different sites. A total of 600 SERS spectra were obtained for PHE and ROS at different concentrations. Then, we employed the multi-tasks deep learning model to carry out qualitative recognition and quantitative analysis from these 600 spectra (480 spectra for training dataset, 120 spectra for testing dataset). The experiments showed that the accuracy for qualitative recognition can reach up to 100%, and the accuracy for quantitative analysis is 99.1%. These results indicate that the proposed multi-tasks deep learning model can provide an effective method to achieve SERS spectroscopy analysis in complex matrix.
Fig. 4. SERS spectra of PHE (a) and ROS (b) with different concentrations (curve i-iii: 1 × 10−4 M, 1 × 10−5 M and 1 × 10−6 M) in the complicated serum matrix. Curve iv shows the control experiment for SERS spectrum of serum itself.
4. Conclusion
In summary, we have successfully designed a single-model multi-tasks deep learning network for simultaneously qualitative recognition and quantitative analysis of SERS spectroscopy. Specifically, the SERS spectra of two kinds of hypoglycemic drugs (PHE, ROS) and the corresponding mixtures were collected, respectively, with the five concentration grades from 10−4 M to 10−8 M. By optimizing the loss functions and other hyperparameters, the best average accuracy can reach up to 98.7%, with qualitative recognition (99.0%) and quantitative analysis (98.4%), respectively. In comparison with the traditional single-model single-task network, the proposed multi-tasks network can simultaneously achieve qualitative recognition and quantitative analysis with high accuracies and reduce the training time, thus expanding the application scenarios of the deep learning model. As a proof of concept, the practical feasibility of this multi-tasks model was demonstrated by using it to achieve qualitative and quantitative analysis of PHE and ROS in complex serum matrix. These results clearly illustrate the potential of our model in qualitative recognition and quantitative analysis of SERS spectroscopy, which provides an algorithmic and experimental basis for efficiently analyzing the large-scale and multiple substances SERS spectroscopy in practical applications, such as food safety, bacteria/viruses analysis and medical diagnostics.
Funding
Beijing Municipal Natural Science Foundation (Z190006); National Key Research and Development Program of China (2021YFA1400800); National Natural Science Foundation of China (11704266, 11774245); National Youth Talent Support Program; Training Program of the Major Research Plan of Capital Normal University; Scientific Research Base Development Program of Beijing Municipal Commission of Education; Beijing Key Laboratory of Metamaterials and Devices.
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. All code regarding the single-model multi-tasks deep learning network is available in [40].
Supplemental document
See Supplement 1 for supporting content.
References
1. A. I. Pérez-Jiménez, D. Lyu, Z. Lu, G. Liu, and B. Ren, “Surface-enhanced Raman spectroscopy: benefits, trade-offs and future developments,” Chem. Sci. 11(18), 4563–4577 (2020). [CrossRef]
2. H. Tian, N. Zhang, L. Tong, and J. Zhang, “In situ quantitative graphene-based surface-enhanced Raman spectroscopy,” Small Methods 1(6), 1700126 (2017). [CrossRef]
3. A.-I. Henry, B. Sharma, M. F. Cardinal, D. Kurouski, and R. P. Van Duyne, “Surface-enhanced Raman spectroscopy biosensing: In vivo diagnostics and multimodal imaging,” Anal. Chem. 88(13), 6638–6647 (2016). [CrossRef]
4. J. F. Li, Y. F. Huang, Y. Ding, Z. L. Yang, S. B. Li, X. S. Zhou, F. R. Fan, W. Zhang, Z. Y. Zhou, D. Y. Wu, B. Ren, Z. L. Wang, and Z. Q. Tian, “Shell-isolated nanoparticle-enhanced Raman spectroscopy,” Nature 464(7287), 392–395 (2010). [CrossRef]
5. S. Yang, X. Dai, B. B. Stogin, and T.-S. Wong, “Ultrasensitive surface-enhanced Raman scattering detection in common fluids,” Proc. Natl. Acad. Sci. U. S. A. 113(2), 268–273 (2016). [CrossRef]
6. E. C. Le Ru and P. G. Etchegoin, “Single-molecule surface-enhanced Raman spectroscopy,” Annu. Rev. Phys. Chem. 63(1), 65–87 (2012). [CrossRef]
7. D.-K. Lim, K.-S. Jeon, H. M. Kim, J.-M. Nam, and Y. D. Suh, “Nanogap-engineerable Raman-active nanodumbbells for single-molecule detection,” Nat. Mater. 9(1), 60–67 (2010). [CrossRef]
8. H. Xu, E. J. Bjerneld, M. Käll, and L. Börjesson, “Spectroscopy of single hemoglobin molecules by surface enhanced Raman scattering,” Phys. Rev. Lett. 83(21), 4357–4360 (1999). [CrossRef]
9. C. Liu, M. Yang, J. Yu, F. Lei, Y. Wei, Q. Peng, C. Li, Z. Li, C. Zhang, and B. Man, “Fast multiphase analysis: Self-separation of mixed solution by a wettability-controlled CuO@Ag SERS substrate and its applications in pollutant detection,” Sens. Actuators, B 307, 127663 (2020). [CrossRef]
10. Z.-W. Yang, L.-Y. Meng, J.-S. Lin, W.-M. Yang, P. Radjenovic, S.-X. Shen, Q.-C. Xu, Z.-L. Yang, Z.-Q. Tian, and J.-F. Li, “3D hotspots platform for plasmon enhanced Raman and second harmonic generation spectroscopies and quantitative analysis,” Adv. Opt. Mater. 7(23), 1901010 (2019). [CrossRef]
11. H. Lu, L. Zhu, C. Zhang, Z. Wang, Y. Lv, K. Chen, and Y. Cui, “Highly uniform SERS-active microchannel on hydrophobic PDMS: a balance of high reproducibility and sensitivity for detection of proteins,” RSC Adv. 7(15), 8771–8778 (2017). [CrossRef]
12. Q. Guo, M. Xu, Y. Yuan, R. Gu, and J. Yao, “Self-assembled large-scale monolayer of Au nanoparticles at the air/water interface used as a SERS substrate,” Langmuir 32(18), 4530–4537 (2016). [CrossRef]
13. D. Sun, G. Qi, S. Xu, and W. Xu, “Construction of highly sensitive surface-enhanced Raman scattering (SERS) nanosensor aimed for the testing of glucose in urine,” RSC Adv. 6(59), 53800–53803 (2016). [CrossRef]
14. C. Zhang, S. Z. Jiang, Y. Y. Huo, A. H. Liu, S. C. Xu, X. Y. Liu, Z. C. Sun, Y. Y. Xu, Z. Li, and B. Y. Man, “SERS detection of R6G based on a novel graphene oxide/silver nanoparticles/silicon pyramid arrays structure,” Opt. Express 23(19), 24811–24821 (2015). [CrossRef]
15. X. Gao, H. Zhang, X. Fan, C. Zhang, Y. Sun, C. Liu, Z. Li, S. Jiang, B. Man, and C. Yang, “Toward the highly sensitive SERS detection of bio-molecules: the formation of a 3D self-assembled structure with a uniform GO mesh between Ag nanoparticles and Au nanoparticles,” Opt. Express 27(18), 25091–25106 (2019). [CrossRef]
16. Z. Li, S. Jiang, Y. Huo, T. Ning, A. Liu, C. Zhang, Y. He, M. Wang, C. Li, and B. Man, “3D silver nanoparticles with multilayer graphene oxide as a spacer for surface enhanced Raman spectroscopy analysis,” Nanoscale 10(13), 5897–5905 (2018). [CrossRef]
17. S.-Y. Liu, X.-D. Tian, Y. Zhang, and J.-F. Li, “Quantitative surface-enhanced Raman spectroscopy through the interface-assisted self-assembly of three-dimensional silver nanorod substrates,” Anal. Chem. 90(12), 7275–7282 (2018). [CrossRef]
18. S.-R. Wu, X.-D. Tian, S.-Y. Liu, Y. Zhang, and J.-F. Li, “Surface-enhanced Raman spectroscopy solution and solid substrates with built-in calibration for quantitative applications,” J. Raman Spectrosc. 49(4), 659–667 (2018). [CrossRef]
19. W. Shen, X. Lin, C. Jiang, C. Li, H. Lin, J. Huang, S. Wang, G. Liu, X. Yan, Q. Zhong, and B. Ren, “Reliable quantitative SERS analysis facilitated by core–shell nanoparticles with embedded internal standards,” Angew. Chem. Int. Ed. 54(25), 7308–7312 (2015). [CrossRef]
20. H. Liu, Z. Yang, L. Meng, Y. Sun, J. Wang, L. Yang, J. Liu, and Z. Tian, “Three-dimensional and time-ordered surface-enhanced Raman scattering hotspot matrix,” J. Am. Chem. Soc. 136(14), 5332–5341 (2014). [CrossRef]
21. G. Krishnan, R. Joshi, T. O’Connor, F. Pla, and B. Javidi, “Human gesture recognition under degraded environments using 3D-integral imaging and deep learning,” Opt. Express 28(13), 19711–19725 (2020). [CrossRef]
22. Z. Wang, H. Zheng, L. Li, J. Liang, X. Wang, B. Lu, Q. Ye, R. Qu, and H. Cai, “Practical multi-class event classification approach for distributed vibration sensing using deep dual path network,” Opt. Express 27(17), 23682–23692 (2019). [CrossRef]
23. S. Ren, K. He, R. Girshick, X. Zhang, and J. Sun, “Object detection networks on convolutional feature maps,” IEEE Trans. Pattern Anal. Mach. Intell. 39(7), 1476–1481 (2017). [CrossRef]
24. C. Szegedy, W. Liu, Y. Jia, S. Pierre, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, “Going deeper with convolutions,” in Conference on Computer Vision and Pattern Recognition (CVPR) (IEEE, 2015), pp. 1–9.
25. X. Fan, W. Ming, H. Zeng, Z. Zhang, and H. Lu, “Deep learning-based component identification for the Raman spectra of mixtures,” Analyst 144(5), 1789–1798 (2019). [CrossRef]
26. W. Lee, A. T. M. Lenferink, C. Otto, and H. L. Offerhaus, “Classifying Raman spectra of extracellular vesicles based on convolutional neural networks for prostate cancer detection,” J. Raman Spectrosc. 51(2), 293–300 (2020). [CrossRef]
27. M. Hino, E. Benami, and N. Brooks, “Machine learning for environmental monitoring,” Nat Sustain 1(10), 583–588 (2018). [CrossRef]
28. H. M. Hao, Y. G. Liang, H. B. Wu, M. L. Bu, and J. H. Huang, “Infrared spectrum recognition method based on symmetrized dot patterns coupled with deep convolutional neural network,” Spectrosc. Spectral Anal. 41, 782–788 (2021).
29. Q. F. Hu and J. Cai, “Research of terahertz time-domain spectral identification based on deep learning,” Spectrosc. Spectral Anal. 41, 94–99 (2021).
30. K. M. He, X. Y. Zhang, S. Q. Ren, J. Sun, and Ieee, “Deep residual learning for image recognition,” in Conference on Computer Vision and Pattern Recognition (CVPR) (IEEE, 2016), pp. 770–778.
31. X. Chen, L. Xie, Y. He, T. Guan, X. Zhou, B. Wang, G. Feng, H. Yu, and Y. Ji, “Fast and accurate decoding of Raman spectra-encoded suspension arrays using deep learning,” Analyst 144(14), 4312–4319 (2019). [CrossRef]
32. J. Zhu, X. Yu, Y. Zheng, J. Li, Y. Wang, Y. Lin, Z. He, W. Zhao, C. Chen, K. Qiu, and J. Wu, “Association of glucose-lowering medications with cardiovascular outcomes: an umbrella review and evidence map,” Lancet Diabetes Endocrinol. 8(3), 192–205 (2020). [CrossRef]
33. S. E. Nissen and K. E. Wolski, “Effect of rosiglitazone on the risk of myocardial infarction and death from cardiovascular causes,” N. Engl. J. Med. 356(24), 2457–2471 (2007). [CrossRef]
34. P. C. Lee and D. Meisel, “Adsorption and surface-enhanced Raman of dyes on silver and gold sols,” J. Phys. Chem. 86(17), 3391–3395 (1982). [CrossRef]
35. S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” in Proceedings of the 32nd International Conference on Machine Learning (ICML, 2015), pp. 448–456.
36. M. Lin, Q. Chen, and S. Yan, “Network in network,” arXiv:1312.4400 (2013).
37. J. Wu, L. Zhang, X. Bu, P. Li, B. Zhao, and Y. Tian, “Determination of the illegal adulteration of natural healthcare products with chemical drugs using surface-enhanced Raman scattering,” Analyst 143(21), 5202–5209 (2018). [CrossRef]
38. L. Zhang, P. Li, X. Bu, J. Wu, X. Zhang, B. Zhao, and Y. Tian, “Combined host-guest complex with coffee-ring effect for constructing ultrasensitive SERS substrate for phenformin hydrochloride detection in healthcare products,” Anal. Bioanal. Chem. 410(29), 7599–7609 (2018). [CrossRef]
39. P. Li, B. Zhou, X. Cao, X. Tang, L. Yang, L. Hu, and J. Liu, “Functionalized acupuncture needle as surface-enhanced resonance raman spectroscopy sensor for rapid and sensitive detection of dopamine in serum and cerebrospinal fluid,” Chem. Eur. J. 23(57), 14278–14285 (2017). [CrossRef]
40. L. Xie, Y. Shen, M. Zhang, Y. Zhong, Y. Lu, L. Yang, and Z. Li, “Code for “Single-model multi-tasks deep learning network for recognition and quantitation of surface-enhanced Raman spectroscopy”,” GitHub (2022), https://github.com/xiely-123/A-single-model-multi-tasks-deep-learning-network-for-recognition-and-quantitation-of-surface-enhance.
