Multiplex Sensing of Complex Mixtures by Machine Vision Analysis of TLC-SERS Images

Abstract

Thin layer chromatography in tandem with surface-enhanced Raman scattering (TLC-SERS) has demonstrated tremendous potentials as a new analytical chemistry tool to detect a wide range of substances from real-world samples. However, it still faces significant challenges of multiplex sensing from complex mixtures due to the imperfect separation by TLC and the resulting interference of SERS detection. In this article, we propose a multiplex sensing method of complex mixtures by machine vision analysis of the scanning image of the TLC-SERS results. Briefly, various pure substances in solution and the complex mixture solution are separated by TLC followed by one-dimensional SERS scanning of the entire TLC plate, which generates TLC-SERS images of all target substances along the chromatography path. After that, a machine vision method is employed to extract the template images from the TLC-SERS images of pure substance solutions. Finally, we apply a feature point matching strategy based on the Winner-take-all principle, which matches the template image of each pure substance with the mixture image to confirm the existence and derive the position of each target substance in the TLC plate, respectively. Our experimental results based on the mixture solution of five different substances show that the proposed machine vision analysis is highly selective, sensitive and does not require artificial analysis of the SERS spectra. Therefore, we envision that the proposed machine vision analysis of the TLC-SERS imaging is an objective, accurate, and efficient method for multiplex sensing of trace level of target substances from complex mixtures.

1. Introduction

Thin layer chromatography in tandem with surface-enhanced Raman scattering (TLC-SERS) has demonstrated enormous potentials as a new analytical chemistry tool and gained increasing research interests in recent years due to its high sensitivity and ease of implementation [1–6]. TLC-SERS synergically integrates multiple functionalities on a single substrate, realizing mixture separation by combining different eluents (mobile phase) with adsorbent layers (stationary phase) and trace level of chemical sensing by acquiring the vibrational spectra of molecules enhanced by plasmonic nanostructures. Successful applications include analyzing many real-world samples for food safety [7, 8], agriculture [9, 10], medical diagnosis [11, 12] and industrial sensing [13–15]. By appropriate choices of the mobile phase and stationary phase, TLC-SERS is very successful at detecting a single target substance from complex chemical and biological samples. For example, Kundan Sivashanmugan et al. separated and detected tetrahydrocannabinol (THC) from saliva using diatomaceous earth based TLC-SERS technology and the experimental results proved its potential in forensic analysis and public health [16]. Dao Thi CamMinh et al. identified sildenafil adulteration in herbal medicines and dietary supplements, and validated herbal products sold on the market [17]. Other TLC-SERS technology applications include detecting Rhodamine B in chili oil [18], Sudan I in chili sauce [19], melamine from real milk samples [20], and histamine in real tuna samples [21].

However, TLC-SERS techniques face significantly higher challenges in detecting multiple target substances due to the imperfect TLC separation. Since the mobile phase is drawn up the stationary phase via capillary action, it will result in poor resolution of multiple target substances, which will induce undesired interferences to the successive SERS measurement. In addition, the large dynamic range of SERS signals from different substances may cause poor signal-to-noise ratios when identifying multiple substances. Nevertheless, some research progress has been made in recent years to detect complex mixtures with large differences of polarities as they have better chromatography resolution. For instance, Zhengdong Shen et al. separated and identified pyrimethanil, pymetrozine, and carbendazim from the mixture and proved the potential of detecting pesticide residues in fruits and vegetables [22]. The same TLC-SERS technique was also applied to detect four polycyclic aromatic hydrocarbons from edible oil samples [23]. Other successful examples include separating and detecting three dyes from dyed wool [24], four chemical substances in adulterated diabetic botanical dietary supplements (BDS) [25], and on-site detection of four types of substituted aromatic contaminants from water [26]. Nevertheless, it is generally difficult to find a suitable eluent to separate arbitrary substances in real-world mixtures. Especially, as the number of mixtures increases, the difficulty of separation becomes much more significant. For example, Hao Li et al. qualitatively analyzed structurally similar adulterants in BDS using TLC-SERS but suffered from the difficulty of separating rosiglitazone acid and pioglitazone hydrochloride [27]. Instead, they had to adopt a two-dimensional correlation spectroscopy to detect the presence of these two substances. Diya Lv et al. faced the same challenge to completely separate ephedrine and pseudoephedrine when detecting ephedrine and its analogues adulterated in slimming dietary supplements [28]. The eventual analysis had to depend on the characteristic peaks of the SERS spectra. Same problems were also reported in detecting other structurally similar substances from complex mixtures [29, 30]. Although manual analysis of the obtained SERS spectra as the fingerprint information can serve as the last choice for some particular cases, this method is difficult to be applied to general mixture analysis due to several intrinsic drawbacks. First, the manual SERS peak analysis only analyzes one or a few characteristic peaks of the target substance and the accuracy cannot be guaranteed when target substances are complex and overlapping. This is especially true considering that many structurally similar substances are both difficult to be separated by TLC and sharing the same types of chemical bonds. Second, as a manual analytical technique to compare the SERS peaks one by one, it strongly depends on the subjective interpretation and capabilities of the operator. As the number of target substances increases, this method becomes unsustainable. Last, the manual SERS peak analysis is time consuming and requires the calibration of each substance at different concentrations, which is impractical to be used as an on-site or point-of-care (POC) sensing technique.

In order to develop an objective, accurate, and efficient TLC-SERS sensing protocol that can detect a wide range of substances from complex mixture, we propose a machine vision method for the analysis of the TLC-SERS image. Compared with traditional manual SERS peak analysis methods, the proposed machine vision method relies on a holistic analysis of the spatial and spectral information of the TLC-SERS image. As the kernel algorithm of the proposed machine vision, the image matching method processes the entire SERS spectra based on the extracted template image rather than the main characteristic SERS peaks, which will greatly improve the accuracy when the SERS peaks of different substances vary drastically. Most importantly, the image matching algorithm can achieve high accuracy even though the target substances have certain overlaps along the TLC path. In addition, we only need to extract the template image of the pure substance from one TLC-SERS experiment and the template image can be applied to detect various concentration mixtures, which greatly improve the expandability of the proposed method. Last but not the least, the machine vision method does not require any manual interference and is fully objective and intelligent, which is highly essential for rapid, on-site detection of complex mixtures.

2. Multiplex TLC-SERS sensing of complex mixtures by machine vision

Figure.1 shows the schematic of the proposed multiplex sensing method by machine vision analysis of the TLC-SERS image. The TLC substrate was fabricated based on diatomite biosilica channel arrays on a glass plate as reported in our previous work [16]. High density silver nanoparticles (Ag NPs) were integrated with the diatomite biosilica arrays by the in-situ growth method [31]. Detailed fabrication processes of the TLC substrate is described in S1 Experiment Section in the Supporting Information (SI). The proposed multiplex sensing method require four steps as shown in Fig.1. First, we conduct the chromatography process of the mixture by spotting 0.5μL sample onto the TLC substrate and use a mixture of methanol, ethyl acetate ether, and ammonia (V/V = 6:2:2) as the mobile phase for 15 min development. Details about the TLC process can be found in our previous work [21]. Second, we obtain the TLC-SERS image through one-dimensional (1-D) scanning along the diatomite biosilica channel using a portable Raman spectrometer, which will collect holistic spatial and spectral information of all target substances. A simple background subtraction method is used to remove fluorescence signals, which will generate a two-dimensional (2-D) image with Raman shift as the horizontal axis and TLC path as the vertical axis. In the third step, the machine vision method is applied to the TLC-SERS image for multiplex sensing based on two distinct processes: template image extraction and image matching analysis. In order to extract template images of various target substances, we need to use pure substances in solution in the TLC-SERS experiment to obtain the TLC-SERS images of pure substances and extract template images of each substance. After that, we will obtain the TLC-SERS image of the mixture sample and perform image matching analysis based on the Winner-take-all principle using previously extracted template images of various pure substances. The final mixture sensing results will be plotted by a visualization image with the distribution of various target substances based on the spatial distribution information of all target substances.

Fig. 1.

Schematic of the multiplex sensing method of complex mixtures by machine vision analysis of the TLC-SERS image

2.1. Acquisition of TLC-SERS images

Traditional TCL-SERS measurement only collects SERS spectra at a few predetermined spots calculated by the retention factor (Rf) of target substances. In this article, we scan the entire 1-D diatomite biosilica channel to generate a TLC-SERS image that contains holistic spatial and spectral information of all target substances. Although more time consuming, the 1-D scanning method provides abundant data for a more comprehensive analysis of the complex mixture sample. In our experiment, pure samples of Histamine (His), Melamine (Mel), Malachite Green (MG), Rhodamine 6G (R6G), and Sudan I with a concentration of 300 ppm were prepared, as well as a mixture sample containing all five substances with the same concentration. We conduct TLC-SERS measurement for each pure substance in solution using a portable Raman spectrometer equipped with a diode laser emitting at 532 nm wavelength. Each SERS spot along the diatomite biosilica channel was illuminated with a laser spot diameter of 100 μm, power of 10 mW, an acquisition time of 1 second, and in the wavenumber range of 410–1800 cm⁻¹, respectively. The 1-D SERS scanning along the diatomite biosilica channel was performed with a distance interval of 100 μm from the beginning of the TLC channel (where the sample is spotted) to the end where the mobile phase stops. The acquired TLC-SERS images have spectral range of 410–1800 cm⁻¹ with 1 cm⁻¹ resolution, and spatial range along the TLC path with 0.1 mm resolution and the length will depend on the chromatography requirement for sufficient chemical separation. Currently, the TLC sample is moved by a manual control stage with precision down to 1 μm, which takes a long time to complete the full scan. An automatic control system will be developed to greatly reduce the signal collection time to less than 15 minutes.

Following the aforementioned method, TLC-SERS images of five pure substances (i.e. 300 ppm of Histamine, Melamine, Malachite Green, Rhodamine 6G, and Sudan I) were obtained separately for the extraction of the template images. The original TLC-SERS images are shown in Fig.S3 in the SI. Using the same procedure, TLC-SERS images of mixtures with the five substances at the concentration of 300 ppm and 1 ppm were obtained as shown in Fig.S4 in the SI. Among the background noise sources, the low-frequency fluorescence signals are the main factor affecting the signal-to-noise ratio [32]. Previous studies have proved that the wavelet transform works well to remove the fluorescent background signals and suppress the noise of SERS signals [33, 34]. The specific process of the wavelet transform adopted in this article for noise suppression is described in S3 section in the SI.

2.2. Template image extraction of pure substances

The objective of extracting the template image from a pure substance image is to find characteristic peaks on the TLC-SERS image by the image processing method. We define the feature point as a point with the largest or lowest pixel value in either direction within a certain range of the image. In the TLC-SERS image, feature points represent certain attributes due to either chromatography or SERS spectra, which can be used to identify different target substances.

The simple noise suppression by the wavelet transform as described in Section 2.1 is not sufficient to extract essential features of target substances. In real TLC-SERS measurement, high-frequency noises coming from the non-uniform TLC plates and the random variation of SERS enhancement factors may induce strong noises that can affect the analysis outcome. For example, Fig.2 (a) is the TLC-SERS image of Sudan I after removing the fluorescent background signals. We can roughly see the distribution position of Sudan I and the SERS spectra. However, there is still a significant amount of high-frequency noises, especially stripe noises along the TLC path on the image. This is probably due to the dissolve of Sudan I in the mobile phase and is subsequently carried and distributed along the entire TLC path. Although machine learning can deal such noises based on a large amount of data training [35], it requires extensive experimental efforts and is generally not preferred for chemical sensing. A more efficient image processing method is preferred to extract the template images of the pure substances. In this section, we will use Sudan I as an example to show the template image extraction process.

Fig. 2.

(a) The TLC-SERS image of Sudan I after removing the fluorescent background signals; (b) The TLC-SERS image after Gaussian blurring; (c) Mask image of the TLC-SERS image based on Fig.2 (b); (d) The TLC-SERS image after Hadamard product of Fig.2 (b) and Fig.2 (c) with marked coordinate of the characteristic peak; (e) The sliding summation curve of the extracted characteristic peak vector; and (f) The TLC-SERS image marked with the location of the template image based on Fig.2 (a).

First, the image is convolved with a Gaussian template to filter out high-frequency interference noises. The image f(x, y) was blurred using Equation (1), where x is the Raman shift, y is the TLC path in the image, h(s, t) is a Gaussian template of m × n, m and n are both odd numbers, a = (m − 1)/2, b = (n − 1)/2, and g(x, y) is the blurred image, the intensity of which was normalized as shown in Fig.2 (b).

$$\text{g}\left(x,y\right)={\sum }_{s=-a}^a{\sum }_{t=-b}^{b}h\left(s,t\right)f\left(x-s,y-t\right)$$ (1)

After the image is blurred, the characteristic peaks are enhanced while the stripe noises are also enhanced at the same time. Since the stripe noises are distributed along the vertical axis, a gradient operation along the vertical axis is performed to generate a mask image that suppresses the stripe noises. The gradient of the Gaussian blurred image g(x, y) along the vertical axis was calculated based on Equation (2) to generate the mask template gy(x, y). Then, the intensity of the mask template was normalized as shown in Fig.2 (c).

$$\text{gy}\left(x,y\right)=\text{g}\left(x,y+1\right)-\text{g}\left(x,y\right)$$ (2)

After that, the Hadamard product operation [36] of the blurred image g(x, y) and the mask image gy(x, y) was performed using Equation (3) to generate the masked image gp(x, y). The ⊙ symbol is the Hadamard product operation.

$$\text{gp}\left(x,y\right)=\text{g}\left(x,y\right)\odot \text{gy}\left(x,y\right)$$ (3)

The masked image gp(x, y) as shown in Fig.2 (d) clearly suppressed the stripe noises and highlighted the characteristic peaks of the pure substance. However, the gradient operation does not have scale invariance, the feature points of the masked image are not the accurate feature points of the original image. On the contrary, the blur operation is not only scale-invariant but also improves the signal-to-noise ratio of the image. Therefore, the feature points detected on the blurred image are more accurate and can represent the feature points on the original image. The maximum value point of the masked image gp(x, y) is identified as the quasi-feature point. Then the coordinate of the quasi-feature point is mapped to the blurred image g(x, y) and the maximum value point in the two-dimensional neighborhood of this coordinate on the blurred image is the feature point. As shown in Fig.2 (d), the feature point is in the white box, and the feature point coordinate is [973, 223]. The abscissa of the feature point is the Raman shift of the characteristic peak of the pure substance, and the ordinate corresponds to the position coordinate of the TLC plate, which can determine the specific position of the chromatographic development of the pure substance.

Finally, a column of the Raman shift value of the TLC-SERS image is extracted according to the abscissa of the feature point to form a characteristic peak vector. The sliding summation is performed on the characteristic peak vector to obtain the sliding summation curve of the characteristic peak. The window position corresponding to the maximum value of the sliding summation curve is the most concentrated position of the pure substance on the TLC plate. Therefore, the window position corresponding to the maximum value of the sliding summation curve is the area of the template image of the pure substance. The size of the sliding summation window is 7 and the sliding summation curve is shown in Fig.2 (e). The position of Sudan I’s template image on the TLC plate is from 220 to 226 (or 22.0–22.6 mm) as shown in Fig.2 (f).

2.3. Imaging matching algorithm based on Winner-take-all principle

The imaging matching algorithm conducts correlation analysis between the template image of a pure substance and the TLC-SERS image of the mixture to find the region that is similar to the template image on the TLC-SERS image of the mixture. In this article, the position of the target substances on the TLC plate is determined by the Winner-take-all principle. As an example to explain the image matching algorithm, we use a mixture containing 300 ppm Histamine, Melamine, Sudan I, Malachite Green, and Rhodamine 6G. The TLC-SERS image of the mixture was preprocessed by the same method for the pure substance in Section 2.1. Fig.3 (a) is the TLC-SERS image of the mixture after removing the fluorescent background signals. In this section, the Sudan I was used as the target substance to show the image matching algorithm based on the Winner-take-all principle.

Fig. 3.

(a) The TLC-SERS image of the mixture with a concentration of 300 ppm after removing the fluorescent background signals; (b) Schematic of the correlation calculation between the extracted template image of Sudan I and the mixture image; (c) Schematic of matching the feature points of the quasi-target region with those of Sudan I’s template image; and (d) The comparison of the ideal matching curve and the original matching curve.

First, we will calculate the matching curve along the TLC path based on the mixture image. The matching curve is calculated by performing correlation between the template image of the pure substance and the TLC-SERS image of the mixture in a row-by-row and top-to-bottom pattern. A schematic of the correlation calculation is shown in Fig.3 (b). In this study, the Pearson product-moment correlation coefficient (PPMCC) [37] is employed as the image correlation calculation method, which measures the strength of the linear relation between two variables x and y. The value of the PPMCC is equal to the covariance between them divided by the product of their respective standard deviations, as shown by Equation (4).

$${\text{r}}_{xy}=\frac{{\sum }_{i=1}^n\left(x_i-\overline{x}\right)\ast \left(y_i-\overline{y}\right)}{\sqrt{{\sum }_{i=1}^n{\left(x_i-\overline{x}\right)}^2}\ast \sqrt{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}$$ (4)

Since the PPMCC is the standardized covariance, it eliminates the influence of the dynamic range of the two variables while still reflecting the degree of similarity of these two variables. This lays the foundation for the matching between TLC-SERS images of the same substance even at different concentrations.

Second, we extract a quasi-target substance region based on the matching coefficient greater than a threshold σ. The quasi-target substance region is then divided into several small regions according to SERS peaks. The two-dimensional maximum of the small region was extracted as the feature point of the quasi-target substance region. The main feature points of the quasi-target substance regions are filtered out according to the feature point intensity. Feature points of the template image of the pure substance were extracted using the same method. By comparing the feature points of the template image and the quasi-target substance region, it is determined whether the quasi-target substance region is the target substance region. For example, if the feature points of the quasi-target substance region are consistent with the feature points of the template image of the pure substance, the matching coefficient is maintained; otherwise, the matching coefficient of the quasi-target substance region is set to 0.9 * σ, where the σ is 0.5. In this way, the gap between the quasi-target substance region and the non-target substance region is increased to facilitate the subsequent ideal matching. Fig.3 (c) is the schematic of the feature points matching of Sudan I’s template image with quasi-target substance regions of the mixture image. In Fig.3 (c), the red triangles are feature points of the quasi-target substance region and the green circles are feature points of Sudan I’s template image. From Fig.3 (c), it is obvious that the abscissa (x-axis) of the feature points of the two images match very well, which indicates that the quasi-target substance region is indeed the target substance region. The original matching curve after feature point matching is shown in Fig.3 (d), which has a high degree of correlation in the range of 12–21 mm.

Finally, an idealized matching is performed on the original matching curve, according to the Winner-take-all principle. All values on the original matching curve were compared with the matching threshold σ. The first and last points on the curve that are greater than the threshold were located. According to the Winner-take-all principle, the matching coefficient between the two points was set to 1 as the target substance region while all other matching coefficient values were set to 0. Such ideal matching curve outlined the distribution of the target substance. The ideal matching curve of Sudan I in the mixture image is shown in Fig.3 (d), which shows that the Sudan I on the TLC plate is from 12.5 mm to 20.8 mm.

3. Results and discussion

3.1. Template extraction of pure target substances

According to the extraction method of template image as described in Section 2.2, the template images of Histamine, Melamine, Malachite Green, and Rhodamine 6G were extracted from TLC-SERS scanning images of pure substances, respectively. The extraction results are shown in Fig.4 (a)~(d).

Fig. 4.

TLC-SERS images marked with the position of the template image of (a) Histamine, (b) Melamine, (c) Malachite Green, and (d) Rhodamine 6G.

3.2. Multiplex TLC-SERS sensing of complex mixture by image matching algorithm based on Winner-take-all principle

According to the imaging matching algorithm based on the Winner-take-all principle as described in Section 2.3, Histamine, Melamine, Malachite Green, and Rhodamine 6G in the mixture with a concentration of 300 ppm were analyzed, respectively. Their matching curves are shown in Fig.5 (a)~(d).

Fig. 5.

Imaging matching curves of (a) Histamine, (b) Melamine, (c) Malachite Green, and (d) Rhodamine 6G.

Fig.6 is a visualization of the multiplex TLC-SERS sensing of complex mixtures. In order to enhance the contrast of the target substance regions, the TLC-SERS image of the mixture as shown in Fig. 6 (a) is divided into three separate target substance regions with different intensity scale bars as shown in Fig.6 (b) to accommodate the large dynamic range of the SERS signal intensity. Table.1 shows the ideal distribution result of the target substances on the TLC plate. From Fig.6 and Table.1, it can be seen that the algorithm can not only match well-separated target substances, but also identify regions where target substances overlap. We would like to point out that the algorithm is based on the holistic analysis of both spectral and spatial information. When target substances have overlapped TLC patterns, the differentiation is indeed based on SERS spectra of each substance, which is incorporated in the image templates. Although we do not need to explicitly compare the SERS signals by manual interference, the image matching process in the algorithm automatically conducted the comparison of SERS signals and return the result in the form of imaging matching curves as shown in Fig.5.

Fig. 6.

(a) The original TLC-SERS image of the mixture marked with the approximate distribution of the target substances; (b) Visualization image of the target substance regions in the TLC-SERS image.

Table. 1.

Distribution of the five target substances in the TLC path

	Rhodamine 6G	Malachite Green	Sudan I	Histamine	Melamine
Target substance distribution in the TLC path (mm)	3.1–6.7	6.2–12.2	12.5–20.8	19.6–29.4	29.3–38.1

3.3. Negative control test

In Section 3.2, the effectiveness of multiplex sensing was confirmed by the positive presence of target substances in the TLC-SERS image. In this section, a negative control test is designed based on the matching with the Sudan I’s template image. Fig.7 (a)~(d) show the matching curves between the TLC-SERS images of four other pure substances with the Sudan I’s template image. It can be seen that the matching curves are all lower than 0.5. This shows that the method has good specificity, which not only accurately identifies the target substance but also has a good inhibition effect on the interference of non-target substances.

Fig. 7.

Comparison of the matching curve between Sudan I’s template image and the TLC-SERS images of the other four pure substances: (a) Histamine, (b) Melamine, (c) Malachite Green, and (d) Rhodamine 6G.

3.4. Sensitivity test

Previous verification tests were based on 300 ppm template images of the pure substances to match the same concentration mixture. To test the sensitivity at lower concentration and verify the capability of imaging matching with different target concentrations, the TLC-SERS image of the mixture with a concentration of 1 ppm was used to match the template images of the pure substances with 300 ppm. First, a mixture sample of the five substances (i.e. Histamine, Melamine, Malachite Green, Rhodamine 6G, and Sudan I) at 1 ppm was prepared. Second, according to the method shown in Fig.1, the mixture sample was separated by TLC using a mobile phase of 1-butanol, ethanol, water (V/V = 4:1:5). Third, the TLC-SERS image of the mixture with 1ppm was acquired following the method as described in Section 2.1 and was preprocessed by the same method used for the mixture with 300 ppm concentration. Finally, the TLC-SERS image of the mixture with 1 ppm was analyzed by the template images of the pure substances with 300 ppm.

Fig.8 (a) is the TLC-SERS image of the mixture marked with the approximate distribution of the target substance. In order to enhance the contrast of the target substance regions, the TLC-SERS image of the mixture is divided into three regions to improve visualization as shown in Fig.8 (b). Table. 2 shows the ideal distribution result of the target substances on the TLC plate. Through the SERS characteristic peaks analysis for the TLC-SERS image of the mixture, it is proved that the distribution of the target substance in the image is consistent with the multiplex sensing result. It shows that this method has high sensitivity, which can not only analyze mixtures of the same concentration, but also identify the mixtures of much lower concentration. In addition, it can be seen that the distributions of the target substances in the two tables are slightly different by comparing Table. 1 and Table. 2, which is possibly due to the different eluents used to separate these two mixtures. The results also verify the robustness of the method by showing the independence to the eluents used in the TLC process.

Fig. 8.

(a) The original TLC-SERS image of the mixture marked with the approximate distribution of the target substances; (b) Visualization image of the target substance regions in the TLC-SERS image.

Table. 2.

Distribution of the five target substances with a concentration of 1 ppm in the TLC path

	Rhodamine 6G	Malachite Green	Sudan I	Histamine	Melamine
Target substance distribution in the TLC path (mm)	40.4–49.7	33.6–38.4	55.2–63.1	24.1–29.3	29.6–33.3

We should point out that one of the main advantages of the proposed machine vision analysis is the capability to handle the large difference in the amplitude of SERS signals, which can be either induced by the difference of substance concentrations, or the Raman scattering efficiency, or both. In this article, the 10~47× difference in SERS signal intensity is induced by the Raman scattering efficiency among various substances. Although we did not conduct experiment with different substance concentrations, the large dynamic range covered by the machine vision analysis proves the feasibility to detect mixed substances with largely different concentrations if they possess similar Raman scattering efficiency.

4. Conclusions

In this article, we propose a multiplex TLC-SERS sensing method of complex mixtures by machine vision analysis of the scanning image. The measurement was conducted based on diatomite biosilica channel array with in-situ growth Ag NPs, which can simultaneously perform TLC separation and SERS sensing on the same platform. Mixture solution are separated by TLC followed by 1-D SERS scanning of the entire TLC plate, which generates TLC-SERS images of all target substances along the chromatography path. Benefited from the comprehensive spatial and spectral information of the TLC-SERS image, the proposed machine vision method conducts a holistic analysis by the kernel algorithm based on the image matching between the extracted template image from pure substances and the mixture TLC-SERS image. Our experimental results of 300 ppm mixtures show that the image matching algorithm can achieve high accuracy even though the target substances have certain overlaps along the TLC path and a large dynamic range of the SERS signals. Negative control testing confirms excellent specificity of the machine vision method. It also shows high sensitivity when the mixture concentration is much lower (1 ppm) than those used to obtain the image templates. Last but not the least, the machine vision analysis does not require any manual interference and is fully objective and intelligent with execution time less than 1 minute on a personal computer, which is highly essential for rapid, on-site detection of complex mixtures. We envision that the proposed machine vision analysis of the TLC-SERS imaging can achieve objective, accurate, and efficient multiplex sensing of trace level of target substances that can play pivotal roles in many applications such as food safety, environmental protection, forensics, and public health.

Research Highlight.

• A multiplex sensing method of complex mixtures by machine vision analysis of the scanning image of the TLC-SERS results

• Holistic analysis by the kernel algorithm based on the image matching between the extracted template image from pure substances and the mixture TLC-SERS image

• Excellent sensitivity down to 1ppm for five mixed chemical substances and high specificity based on negative control testing

• No requirement of any manual interference and fully objective, intelligent, rapid without excessive training

Biography

Alan X. Wang is an Associate Professor of the School of Electrical Engineering and Computer Science at Oregon State University since 2011. He received his Ph.D. degree in Electrical and Computer Engineering from the University of Texas at Austin in 2006. From 2007–2011, he was with Omega Optics, Inc., where he served as the Chief Research Scientist for 9 SBIR/STTR projects. His research interests include nanophotonic devices for optical interconnects, and optical sensors for chemical and biological detection. His current research activities are sponsored by the National Science Foundation, the National Institutes of Health, Oregon Nanoscience and Microtechnologies Institute, the National Energy Technology Laboratory, and industrial sponsors such as Hewlett-Packard. He has more than 90 journal publications and 100 conference presentations, and also holds five U.S. patents. He is a senior member of IEEE Photonics, SPIE and OSA.