Versatile band-pass filters for fluorescence imaging of the food products for quality assessment

Abstract

Determination of food quality using fluorescence measurements has been achieved with high accuracy in many previous studies. However, the use of fluorescence measurements for determining the quality of food products and authenticating them remains limited due to the high cost involved. In this study, we propose versatile band-pass filters for fluorescence imaging of food products for quality assessment by simulation. The results showed that the proposed band-pass filters have similar accuracies to previously reported methods and are practical in most cases such as estimating aflatoxin contamination in nutmeg, inosine 5′-monophosphate (IMP) in frozen fish, and the geographical origin of mangos. Furthermore, these filters can reduce the number of variables in the prediction model and are thereby expected to reduce the measurement time and filter cost when applied to actual imaging systems.

Introduction

In recent decades, public interest in food quality and production has increased. This is likely related to changes in eating habits, consumer behavior, and the development and increased industrialization of food supply chains. The demand for high quality and safety in food production calls for high standards for quality and process controls, which requires sensitive and rapid analytical tools to assess food (Christensen et al., 2006). An excitation-emission matrix (EEM) (Warner et al., 1977), also known as a fluorescence fingerprint (Trivittayasil et al., 2018), is a set of fluorescence spectra acquired at consecutive excitation wavelengths to create a three-dimensional diagram. The EEM has been widely applied for the nondestructive measurement of the physical and chemical properties of objects (Sorrell et al., 1994; Wolf et al., 2001; Sterenborg et al., 1994; Werkhaven et al., 1994). In food product quality assessments, the EEM can determine several properties (functional, composition, degree of nutrition, and origin) of animal (e.g., dairy, meat, fish, and egg) and plant (oils, cereal, sugar, fruit, and vegetable) products, and can also identify bacteria of agro-alimentary interest without the use of chemical reagents (Karoui and Blecker, 2011). In previous studies, various methods were proposed to improve the estimation accuracy of this technique, such as using a fixed excitation wavelength (Karoui et al., 2006a; Karoui et al., 2006b; Olsen et al., 2006), using specific excitation-emission combinations (Hasegawa et al., 1992; Dufour et al., 2003; Trivittayasil et al., 2018), or using the ratio of the fluorescence intensity for two specific excitation-emission combinations (Birlouez et al., 1998; Aubourg et al., 1998) instead of using the whole EEM dataset.

Determination of chemical properties or classification of the geographical origin of food can be carried out with high accuracy using EEM. However, adopting fluorescence as a technique for determining quality and authenticating food products remains limited because most EEM-based methods are point measurements and not suitable for large targets or mass inspection. In order to solve this problem, hyperspectral imaging was applied using multiple excitation-emission combinations to visualize the distribution of target properties instead of performing point estimation (Nishino et al., 2013; Bui et al., 2018). A schematic diagram of a fluorescence imaging system is shown in Fig. 1. In fluorescence imaging, the sample is illuminated by a light source in a specific excitation wavelength range, which is limited by band-pass filters. The fluorescence images of the sample are then measured by a camera equipped with band-pass filters. Each fluorescence image corresponds to one excitation band-pass filter and emission band-pass filter combination and is represented by one rectangle window in the EEM (Fig. 1). Finally, sample properties such as compound concentrations or the geographical origin are estimated from the obtained fluorescence images. Even though fluorescence imaging can measure an entire sample and multiple samples simultaneously, challenges still exist regarding the implementation of this technology at the food industry level. Excitation-emission band-pass filter combinations on the light-source and camera sides are target dependent, which means that different wavelengths are chosen for different estimation targets at the calibration stage. Therefore, different band-pass filters are needed for different target parameters, making it a costly technique.

Fig. 1.

Schematic diagram of the fluorescence imaging system (left) and the areas corresponding to fluorescence images in EEM (right).

This study was conducted to find versatile excitation-emission band-pass filters for fluorescence imaging of different food products for quality assessment. The present study was conducted in two phases. In the first phase, we selected 70 compounds related to food nutrition, freshness, and umami components as samples for fluorescence spectra (EEM) measurement. From the obtained EEM, we generated a synthetic EEM dataset for analysis. Parallel factor analysis (PARAFAC) was applied to the generated synthetic EEM dataset in order to determine the excitation-emission wavelengths for the band-pass filters. In the second phase, the practicality of the proposed band-pass filters was investigated by employing them to a real problem.

Materials and Methods

Sample selection    In this study, compounds related to food nutrition (e.g., vitamins and essential amino acids), freshness, and umami components were selected as samples for fluorescence spectral measurements. Furthermore, compounds reported in previous studies, such as chlorophyll and ferulic acid, were also selected for the experiment (Bron et al., 2004; Codrea et al., 2004; Ram et al., 2004). Moreover, a list of the top 100 compounds in food was compiled from FooDB version 1.0, the world's largest and most comprehensive resource on food constituents, chemistry, and biology (www.foodb.ca). Then, only compounds with a conjugated system were selected as experimental samples since they have the capability to produce fluorescence. In total, 72 compounds were selected as samples for fluorescence spectral measurements. All reagents used in the experiments were of special grade or higher (see Appendix Table S1).

Table S1. List of compounds used in the experiment.

No.	CAS No.	Name	Company	Fluorescence	Solvent	Concentration	Photomultiplier
						(ppm)	(V)
1	10191-41-0	(±)-α-Tocopherol	Wako Pure Chemical Inds. Ltd.	yes	ethanol	10	600
2	16178-48-6	ADP disodium salt	Oriental Yeast Co., Ltd.	yes	phosphate	100	100
3	18422-05-4	AMP	Oriental Yeast Co., Ltd.	yes	Mili-Q	100	900
4	520-36-5	Apigenin	Wako Pure Chemical Inds. Ltd.	none
5	51963-61-2	ATP disodium salt hydrate	Oriental Yeast Co., Ltd.	yes	Mili-Q	100	800
6	58-85-5	Biotin	Sigma-Aldrich	none
7	331-39-5	Caffeic acid	Wako Pure Chemical Inds. Ltd.	yes	ethanol	100	800
				yes	phosphate	100	900
8	1958/8/2	Caffeine	Wako Pure Chemical Inds. Ltd.	yes	phosphate	100	900
9	327-97-9	Chlorogenic acid	Wako Pure Chemical Inds. Ltd.	yes	ethanol	100	800
10	479-61-8	Chlorophyll a	Wako Pure Chemical Inds. Ltd.	yes	ethanol	10	500
11	519-62-0	Chlorophyll b	Sigma-Aldrich	yes	ethanol	10	600
12	67-97-0	Cholecalciferol	Wako Pure Chemical Inds. Ltd.	none
13	68-19-9	Cyanocobalamin	Sigma-Aldrich	none
14	486-66-8	Daidzein	Wako Pure Chemical Inds. Ltd.	yes	phosphate	100	800
15	552-66-9	Daidzin	Wako Pure Chemical Inds. Ltd.	none
16	6893-26-1	D-Glutamic acid	Wako Pure Chemical Inds. Ltd.	none
17	6217-54-5	Doconexent	Sigma-Aldrich	yes	ethanol	100	800
18	10417-94-4	Eicosapentaenoic acid	Sigma-Aldrich	yes	ethanol	100	900
19	50-14-6	Ergocalciferol	Wako Pure Chemical Inds. Ltd.	none
20	1135-24-6	Ferulic acid	Wako Pure Chemical Inds. Ltd.	yes	phosphate	100	800
21	59-30-3	Folic aid	Wako Pure Chemical Inds. Ltd.	yes	phosphate	100	900
22	485-72-3	Formononetin	Sigma-Aldrich	yes	ethanol	10	700
23	446-72-0	Genistein	Wako Pure Chemical Inds. Ltd.	none
24	529-59-9	Genistin	Sigma-Aldrich	none
25	617-65-2	Glutamic acid	Wako Pure Chemical Inds. Ltd.	none
26	85-32-5	Guanylic acid	Combi-Blocks	none
27	68-94-0	Hypoxanthine	Wako Pure Chemical Inds. Ltd.	none
28	131-99-7	IMP	Junsei Chemical Co. Ltd.	yes	Mili-Q	100	900
29	120-72-9	Indole	Wako Pure Chemical Inds. Ltd.	yes	phosphate	100	500
30	58-63-9	Inosine	Junsei Chemical Co. Ltd.	none
31	520-18-3	Kaempferol	Wako Pure Chemical Inds. Ltd.	none
32	50-81-7	L(+)-Ascorbic acid	Wako Pure Chemical Inds. Ltd.	none
33	61-90-5	Leucine	Wako Pure Chemical Inds. Ltd.	none
34	56-86-0	L-Glutamic acid	Wako Pure Chemical Inds. Ltd.	none
35	71-00-1	L-Histidine	Wako Pure Chemical Inds. Ltd.	none
36	73-32-5	L-Isoleucine	Wako Pure Chemical Inds. Ltd.	none
37	63-68-3	L-Methionine	Wako Pure Chemical Inds. Ltd.	none
38	147-85-3	L-Proline	Wako Pure Chemical Inds. Ltd.	none
39	73-22-3	L-Tryptophan	Wako Pure Chemical Inds. Ltd.	yes	phosphate	10	600
40	491-70-3	Luteolin	Wako Pure Chemical Inds. Ltd.	none
41	502-65-8	Lycopene	Wako Pure Chemical Inds. Ltd.	yes	acetonitrile	1	900
42	56-87-1	Lysine	Wako Pure Chemical Inds. Ltd.	none
43	580-72-3	Matairesinol	Sigma-Aldrich	yes	ethanol	10	600
44	58-27-5	Menadione	Wako Pure Chemical Inds. Ltd.	yes	ethanol	100	800
45	529-44-2	Myricetin	Wako Pure Chemical Inds. Ltd.	none
46	606-68-8	NADH	Oriental Yeast Co., Ltd.	yes	phosphate	100	800
47	2646-71-1	NADPH	Combi-Blocks	yes	phosphate	100	800
48	98-92-0	Niacinamide	Wako Pure Chemical Inds. Ltd.	none
49	59-67-6	Nicotinic acid	Wako Pure Chemical Inds. Ltd.	yes	ethanol	100	900
50	63-91-2	Phenylalanine	Peptide Institute, Inc.	yes	phosphate	100	800
51	84-80-0	Phytonadione	Wako Pure Chemical Inds. Ltd.	yes	ethanol	100	900
52	487-36-5	Pinoresinol	Sigma-Aldrich	yes	acetonitrile	10	500
53	99-50-3	Protocatechuic acid	Wako Pure Chemical Inds. Ltd.	yes	ethanol	10	600
54	65-23-6	Pyridoxine	Sigma-Aldrich	yes	ethanol	10	600
				yes	phosphate	10	600
55	120-80-9	Pyrocatechol	Wako Pure Chemical Inds. Ltd.	yes	phosphate	10	700
56	68-26-8	Retinol	Sigma-Aldrich	yes	ethanol	10	800
57	127-47-9	Retinol acetate	Wako Pure Chemical Inds. Ltd.	yes	ethanol	10	800
58	83-88-5	Riboflavine	Sigma-Aldrich	yes	phosphate	100	600
59	530-59-6	Sinapinic acid	Sigma-Aldrich	yes	phosphate	100	900
60	83-67-0	Theobromine	Wako Pure Chemical Inds. Ltd.	none
61	24539	Thiamine hydrochloride	Wako Pure Chemical Inds. Ltd.	none
62	72-19-5	Threonine	Wako Pure Chemical Inds. Ltd.	none
63	60-18-4	Tyrosine	Wako Pure Chemical Inds. Ltd.	yes	phosphate	10	600
64	501-94-0	Tyrosol	Wako Pure Chemical Inds. Ltd.	yes	phosphate	10	600
65	72-18-4	Valine	Wako Pure Chemical Inds. Ltd.	none
66	121-33-5	Vanillin	Wako Pure Chemical Inds. Ltd.	yes	ethanol	100	900
67	137-08-6	Vitamin B5	Wako Pure Chemical Inds. Ltd.	none
68	144-68-3	Zeaxanthin	Sigma-Aldrich	none
69	7488-99-5	α-Carotene	Wako Pure Chemical Inds. Ltd.	yes	ethanol	10	900
70	472-70-8	β-Cryptoxanthin	Wako Pure Chemical Inds. Ltd.	yes	acetonitrile	1	900
71	53-84-9	β-NAD+	Oriental Yeast Co., Ltd.	yes	Mili-Q	100	900
72	1184-16-3	β-NADP+	Oriental Yeast Co., Ltd.	yes	Mili-Q	100	900

Fluorescence spectral measurement (EEM)    For EEM measurements, Milli-Q water, phosphate buffer, ethanol, and acetonitrile were used as solvents. The reagents were dissolved in solvents at concentrations of 0.1, 1, 10, and 100 ppm (1 ppm = 1 mg/L). Then, 300 µL of each dissolved sample was placed into the SQ grade quartz micro cuvette (GL Sciences FM20-SQ-3). EEM was measured using a fluorescence spectrophotometer (F-7000, Hitachi High-Tech Science Corporation) with a right-angle geometry. The cuvette was placed in the spectrophotometer with the shorter path length side facing the excitation light source. The excitation wavelength was changed at 5 mm intervals in the range 200–800 nm, and the emission wavelength was detected at 2 mm intervals in the range 200–800 nm. The slit width on both sides was 5 nm. The photomultiplier voltage was adjusted between 400 and 900 V with a scanning speed of 30 000 nm/min. The solution concentrations and the photomultiplier voltage were adjusted to acquire an EEM with the lowest signal-to-noise ratio.

The raw EEM data includes some non-fluorescence components. There is hypothetically no emission below the excitation wavelength based on Stokes' shift. Further, owing to light scattering effects such as Raman and Rayleigh scattering, a wavelength region where fluorescence and scattered light are superimposed typically exists in any EEM (Rinnan and Andersen, 2005; Andersen and Bro, 2003). Therefore, the scattering signals and areas with emission wavelengths that are shorter than excitation wavelengths do not carry relevant chemical information and were excluded from the EEM.

Based on the detected fluorescence, standard reagents were selected. The EEM for the standard reagents contains the fluorescence spectra of the corresponding compounds. However, food contains a large variety of compounds with overlapping fluorescence spectra. For this reason, in this study, the EEMs for the standard reagents were used to generate synthetic EEM data, which were the random weighted sum of EEMs for the standard reagents with a uniform distribution, and with the sum of the weights normalized to 1. In total, one thousand synthetic EEM spectra were generated.

Parallel factor analysis    The obtained EEM data is three-way data (excitation, emission, and samples), which provides a large amount of information. PARAFAC is a multivariate analysis method originating from psychometrics (Harshman et al., 1970; Carroll et al., 1970). In contrast to conventional methods such as principal component analysis (PCA), which deals with two-way array data (variable and sample axes), PARAFAC is a method for separating components from multi-way data. In EEM data analysis, PARAFAC separates overlapping fluorescence spectra into individual spectra corresponding to component chemicals in the sample. The results of PARAFAC are a set of scores and loadings, which is the same as for PCA. However, in contrast to PCA, PARAFAC yields two sets of loadings, which are emission loadings and excitation loadings.

A major practical obstacle for using the PARAFAC model is the need to determine the appropriate number of components. The number of components in the PARAFAC model can be obtained from the core consistency diagnostic, which evaluates the ‘appropriateness’ of the model (Bro and Kiers, 2003). The core consistency is always less than or equal to 100% and may also be negative. A core consistency close to 100% implies an appropriate model. If a dataset is modeled using PARAFAC with an increasing number of components, the core consistency will typically decrease more or less monotonically and gradually, and then will decrease abruptly beyond a certain number of components (Bro and Kiers, 2003).

In practice, for real-world non-ideal datasets, the core consistency is not always a reliable diagnostic for finding the required number of PARAFAC components (Murphy et al., 2013). Harshman proposed a method called split-half analysis to confirm that the PARAFAC model is appropriate (Harshman and Lundy, 1994). The split-half analysis examines different subsets of the data independently, and the same result (same loadings) will be obtained in the non-split modes from models of any suitable subset of the data if the correct number of components is chosen. If too few or too many components are chosen, the model parameters will differ if the model is fitted to different datasets.

In this study, PARAFAC and split-half analysis were applied to extract the principle components from the generated EEM dataset. Core consistency and similarity of split-half analysis were used as indicators for determining the appropriate number of components. Savitzky-Golay smoothing (Savitzky and Golay, 1964) along with emission wavelength axis and data normalization, in which the fluorescence intensity at each wavelength is divided by the summed squared value of the intensities at all wavelengths (i.e., normalized to have the same multivariate vector length), were applied before the analysis. PLS_Toolbox v8.8.1 (Eigenvector Inc. Wenatchee, USA) was used for the PARAFAC and split-half analyses. Subsequently, the excitation-emission band-pass filters were determined from the principle component loadings.

Verifying proposed band-pass filters    Three EEM datasets related to food from previous studies were used for verification (Aiyama et al., 2018; Bui et al., 2018; Nakamura et al., 2012). A summary of the parameters for these datasets is shown in Table 1. PLS_Toolbox v8.8.1 was used for the analysis. The analysis methods and calibration vs. validation ratio were the same or as similar as possible to those in previous studies (Aiyama et al., 2018; Bui et al., 2018; Nakamura et al., 2012). The 4-filter and 7-filter datasets were calculated by summing the fluorescence intensities of the corresponding wavelength ranges in the original EEM data. Accordingly, four and seven variables corresponding to 4-filter and 7-filter datasets were obtained. Subsequently, four types of preprocessing methods (mean center, autoscale, normalize + mean center, normalize + autoscale) were applied to the filter data, and the one with the best results was selected as the optimum preprocessing method.

Table 1. Details of data-sets from previous studies: Nutmeg, Frozen fish, and Mango

Data-set	Nutmeg (Aiyama et al., 2018)	Frozen fish (Bui et al., 2018)	Mango (Nakamura et al., 2012)
Ex	250–700 nm, 10 nm step	250–800 nm, 10 nm step	200–870 nm, 10 nm step
Em	260–720 nm, 10 nm step	250–800 nm, 10 nm step	230–900 nm, 10 nm step
Sample	calibration 61, validation 30	48	544
Target	aflatoxin contamination	K-value, IMP	geographic origin
Previous method	PLS	PLS	canonical discriminant analysis
Applied method	PLS	PLS	PLSDA

Results and Discussion

EEM preprocessing and generation of synthetic EEM data    Of the 70 reagents tested, 41 reagents emitted fluorescence, and two of these (caffeic acid and pyridoxine) emitted fluorescence in two different solvents (see Appendix Table S1). Therefore, 43 EEM showing fluorescence were obtained. Fig. 2(a) shows the EEM spectra of adenosine 5′-triphosphate (ATP) obtained from fluorescence spectral measurements. The preprocessed EEM spectra masked only the fluorescence area after removing the irrelevant areas as shown in Fig. 2(b). Those reagents were selected as standard reagents, and the preprocessed EEM spectra were used to generate a synthetic EEM dataset. In total, 1 000 synthetic EEM spectra were generated by synthesizing 43 EEMs for standard reagents with random weights. One representative generated EEM spectra is shown in Fig. 2(c).

Fig. 2.

EEM preprocessing to remove areas that contains emission wavelengths shorter than the excitation wavelengths and scattering effect.

(a) Raw EEM data obtained from measurement. (b) EEM data after preprocessing. (c) One example of generated synthetic EEM.

Defining excitation-emission band-pass filters    The results of applying PARAFAC to the generated synthetic EEM dataset for different component numbers are shown in Fig. 3(a). In this study, a split-half analysis was also applied to analyze the generated synthetic EEM dataset for different component numbers. The results are shown in Fig. 3(b). In Fig. 3(a), for one to three components, the core consistency was close to 100%. The core consistency decreased with increasing number of components and became negative for nine components or more. This is because the influence of noise and other non-trilinear variations increases with increasing number of components. In Fig. 3(b), for one to four and for seven components, the similarity of the results (loadings) from the split-half analysis was high (more than 70%).

Fig. 3.

Core consistency of PARAFAC model (a), and similarity of split-half model (b) at different component number.

PARAFAC and split-half analysis were applied to determine the appropriate number of excitation-emission band-pass filters and not for separating components from the generated synthetic EEM dataset. Therefore, four and seven were selected as the appropriate number of excitation-emission band-pass filters. PARAFAC with four and seven components was applied to a generated synthetic EEM dataset, and the obtained loadings are shown in Figs. 4 and 5, respectively. For each loading, areas containing scattering signals and areas where the emission wavelength was shorter than the excitation wavelength were excluded (same as Fig. 2(b)). Since this study aimed to determine band-pass filters for fluorescence imaging, which are excitation-emission wavelength combinations as indicated by the rectangle windows in Fig. 1, all possible windows within the preprocessed range for each loading were considered. Subsequently, the sum of the loadings covered by each window was calculated. The window covering the largest sum of loadings was defined as the excitation-emission band. The excitation-emission bands for four and seven components are shown as purple rectangles in Fig. 4, Fig. 5, and Table 2. Components for the 4-filter and 7-filter sets were taken as candidates for the versatile filter.

Fig. 4.

Loadings of four components PARAFAC and defined excitation-emission bands of 4-filter (purple rectangles).

Fig. 5.

Loadings of seven components PARAFAC and defined excitation-emission bands of 7-filter (purple rectangles).

Table 2. Wavelength range of proposed 4-filter and 7-filter bands

No.	4-filter		7-filter
	Ex [nm]	Em [nm]	Ex [nm]	Em [nm]
1	265–375	396–504	220–300	322–414
2	215–295	316–404	300–405	426–574
3	320–475	496–614	260–385	406–494
4	380–615	636–734	295–360	382–564
5			330–485	506–634
6			200–245	266–374
7			380–615	636–734

The results show that filters 1, 2, 3, and 4 in the 4-filter set are the same or similar to filters 3, 1, 5, and 7 in the 7-filter set. Each filter seems to represent common fluorescence peaks of several compounds. The first filter covers the fluorescence peaks for lycopene, NADH, NADPH, some vitamins (folic acid, menadione, and phytonadione), isoflavones (daidzein and formononetin), chlorogenic acid, eicosapentaenoic acid, ferulic acid, sinapinic acid, vanillin, and α-carotene. The second filter mainly covers the fluorescence peaks for ATP-related compounds (AMP, ADP, ATP, IMP, β-NAD+, and β-NADP+), amino acid (tryptophan, tyrosine, and phenylalanine), phenols (tyrosol and pyrocatechol), some vitamins (α-tocopherol, nicotinic acid, and pyridoxine), lignan (matairesinol and piroresinol), indole, and protocatechuic acid. The third filter mainly covers the fluorescence peaks for some vitamins (retinol, retinol acetate, riboflavin, and phytonadione). The last filter mainly covers the fluorescence peaks for some carotenoids (β-cryptoxanthin, chlorophyll a, and chlorophyll b), and doconexent.

The remaining filters in the 7-filter set are filters 2, 4, and 6, which are different from the 4-filter set. Filters 2 and 4 overlap parts of filters 3 and 5 and therefore cover the fluorescence peaks for the same compounds. In addition, filter 2 covers the fluorescence peaks for caffeic acid (ethanol and phosphate solvent), and filter 4 covers the fluorescence peaks for caffeine. Filter 6 is close to filter 1 and therefore covers the fluorescence peaks for amino acid (part of tryptophan, tyrosine, and phenylalanine), phenols (tyrosol and pyrocatechol), and some vitamins (α-tocopherol and part of pyridoxine), lignan (matairesinol and piroresinol), and indole.

Verifying the proposed band-pass filters    The results of applying the band-pass filters to datasets for previous studies are shown in Table 3. In this study, the EEM for the reagents was measured at 5 nm intervals for excitation wavelengths and 2 nm intervals for emission wavelengths. The results showed that the 4-filter and 7-filter sets have 5 nm intervals for excitation wavelengths and 2 nm intervals for emission wavelengths. However, the EEM datasets from previous studies were measured with 10 nm intervals (Aiyama et al., 2018; Bui et al., 2018; Nakamura et al., 2012). Therefore, in this analysis, the wavelength range for the 4-filter and 7-filter sets were rounded to the nearest and larger wavelength range to cover the whole defined band-pass area. For example, filter 1 in the 4-filter for Ex = 265–375 nm, Em = 396–504 nm was rounded to Ex = 260–380 nm, Em = 390–510 nm in the 10 nm interval dataset. The whole EEM, which means the entire EEM data after removing the scattering signals and non-fluorescence areas of the datasets (Aiyama et al., 2018; Bui et al., 2018; Nakamura et al., 2012), was used for analysis.

Table 3. Comparison between previous studies and proposed band-pass filters based on model accuracy and number of variables

Data-set	Nutmeg	Frozen fish		Mango
	Aflatoxin contamination	K-value	IMP	Geographic origin
	R2	R2	R2	Classification accuracy -%
Previous study	0.69 (whole EEM*, 853)	0.78 (26)	0.82 (26)	92.30 (14)
Whole EEM*		0.91 (1 054)	0.87 (1 054)	91.92 (2 063)
4-filter	0.69 (4)	0.30 (4)	0.78 (4)	70.69 (4)
7-filter	0.71 (7)	0.56 (7)	0.84 (7)	82.20 (7)

*  Analysis results using whole EEM data after removal of the scattering signals and non-fluorescence area.

**  Numbers in ( ) along with R² or classification accuracy indicate the number of variables used.

For the nutmeg dataset, the coefficient of determination (R²) for predicting aflatoxin contamination was 0.69 when using the 4-filter set and 0.71 when using the 7-filter set, which is similar to the result obtained in a previous study (0.69) (Aiyama et al., 2018). However, in the previous study, a whole EEM with 853 variables was used for prediction, and the number of variables was significantly higher than for the band-pass filters with four and seven variables used in the present study. According to the previous study (Aiyama et al., 2018), the variable importance in projection (VIP) value for the aflatoxin concentrations prediction model showed some peaks at Ex = 250, 320, 390, 460, and 520 nm, and Em = 420, 420, 490, 720, and 640 nm, where Ex_{250 nm}/Em_{420 nm} reflects the fluorescence of aflatoxin and Ex_{390 nm} /Em_{490 nm} reflects the fluorescence of kojic acid derivatives. These five excitation-emission areas are all covered or are very close to the wavelength range of the 4-filter and 7-filter sets. This explains why both types of filter showed similar performance as in the previous study (Aiyama et al., 2018).

For the frozen fish dataset, in predicting inosine 5′-monophosphate (IMP), the coefficient of determination of prediction was 0.78 when using the 4-filter set and 0.84 when using the 7-filter set, which is similar to the result of the previous study (0.82) (Bui et al., 2018). In predicting the K-value, which is the freshness index of fish, the coefficient of determination for the prediction was 0.30 when using the 4-filter set and 0.56 when using the 7-filter set, which are considerably worse than the results of the previous study (0.78) (Bui et al., 2018). The K-value is calculated using the concentrations of six compounds, that is, adenosine 5′-triphosphate (ATP), 5′-diphosphate (ADP), adenosine 5′-monophosphate (AMP), IMP, inosine (HxR), and hypoxanthine (Hx) (Saito et al., 1959), where only ATP, ADP, AMP, and IMP emit fluorescence. In a previous study (Bui et al., 2018), the K-value was estimated by using the fluorescence image obtained at Ex = 340 nm, Em = 380–630 nm with a 10 nm interval. In the study by ElMasry et al. (2015), the K-value was estimated using 11 excitation-emission combinations (Ex = 480 nm; Em = 550, 590, 610, 670, 700, 710, 730, 750, and 770 nm). In the present study, the fluorescence peak areas for ATP, ADP, AMP, and IMP were only covered by filters 1 and 2 in the 4-filter set, and filters 1 and 3 in the 7-filter set, which might have restricted the resolution. Hence, the predicted K-values were less accurate than those obtained in the previous study (Bui et al., 2018).

For the mango dataset, three localities in different geographical regions (Taiwan, Miyazaki, and Okinawa) were used as classification targets. The classification accuracy was 70.69% when using the 4-filter set and 82.20% when using the 7-filter set. These accuracy values were less than those using the whole EEM (91.92%) and a previous study (92.30%) (Nakamura et al., 2012), but this is a reasonable accuracy for such a small number of filters. There were 14 combinations of excitation-emission wavelengths used for classification in the previous study, and the number of variables is significantly larger than the defined band-pass filters. In a previous study (Nakamura et al., 2012), it was suggested that the fluorescence information in the wavelength range of Ex = 260–290 nm and Em = 340–360 nm contributed significantly to the determination of the geographic origin for mangoes. This wavelength range was covered by filter 2 in the 4-filter set and filter 1 in the 7-filter set, which suggests that these filter sets worked with reasonable accuracy for a smaller number of variables.

These results show that the proposed band-pass filters have similar accuracy to previously reported methods (Aiyama et al., 2018; Bui et al., 2018; Nakamura et al., 2012). The prediction and classification accuracies are higher for the 7-filter set than for the 4-filter set. Moreover, the filters were able to reduce the number of variables in the prediction model to four and seven, respectively, thereby reducing the measurement time and filter cost compared to applying the same wavelength combinations as in previous studies (Aiyama et al., 2018; Bui et al., 2018; Nakamura et al., 2012) to the fluorescence imaging system. This suggests that the proposed filters could be used as versatile filters for fluorescence imaging of food products for quality assessment.

Conclusion

This study proposed novel versatile band-pass filters for fluorescence imaging of food products for quality assessment. A practical level of accuracy was achieved in most cases, even though the number of auto-fluorescent compounds used in filter design was limited. The accuracy was acceptable in most cases even though the number of variables in the prediction model was reduced. If such filters are employed for different targets, mass production and cost reductions, as have been achieved for filters for RGB cameras, can be expected. Therefore, this approach offers a more practical way of adopting fluorescence measurements for determining quality and authenticating food products than using target-specific filters. In this study, however, the performance of the proposed band-pass filters was not tested using actual instruments, but only through simulations. This means that the camera sensitivity, filter transmittance, and environmental noise were not considered. Experimental validation of the performance is a subject for future work.

Acknowledgements    This work was supported by the Leading Graduate School Program R03 of MEXT.

References

•  Aiyama,  R.,  Trivittayasil,  V., and  Tsuta,  M. (2018). Discrimination of aflatoxin contamination level in nutmeg by fluorescence fingerprint measurement. Food control, 85, 113-118.
•  Andersen,  C. M. and  Bro,  R. (2003). Practical aspects of PARAFAC modeling of fluorescence excitation-emission data. J. Chemom., 17, 200-215.
•  Aubourg,  S. P.,  Sotelo,  C. G., and  Pérez-Martín,  R. (1998). Assessment of quality changes in frozen sardine (Sardina pilchardus) by fluorescence detection. J. Am. Oil Chem.' Soc., 75, 575-580.
•  Birlouez-Aragon,  I.,  Nicolas,  M.,  Metais,  A.,  Marchond,  N.,  Grenier,  J., and  Calvo,  D. (1998). A rapid fluorimetric method to estimate the heat treatment of liquid milk. Int. Dairy J., 8, 771-777.
•  Bro,  R. and  Kiers,  H. A. L. (2003). A new efficient method for determining the number of components in PARAFAC models. J. Chemom., 17, 274-286.
•  Bron,  I. U.,  Ribeiro,  R. V.,  Azzolini,  M.,  Jacomino,  A. P., and  Machado,  E. C. (2004). Chlorophyll fluorescence as a tool to evaluate the ripening of ‘Golden’ papaya fruit. Postharvest Biol. Technol., 33, 163-173.
•  Bui,  M. V.,  Rahman,  M. M.,  Nakazawa,  N.,  Okazaki,  E., and  Nakauchi,  S. (2018). Visualize the quality of frozen fish using fluorescence imaging aided with excitation-emission matrix. Opt. express, 26, 22954-22964.
•  Christensen,  J.,  Nørgaard,  L.,  Bro,  R., and  Engelsen,  S. B. (2006). Multivariate autofluorescence of intact food systems. Chem. Rev., 106, 1979-1994.
•  Codrea,  M. C.,  Nevalainen,  O. S.,  Tyystjärvi,  E.,  Vandeven,  M., and  Valcke,  R. (2004). Classifying apples by the means of fluorescence imaging. Int. J. Pattern Recognit. Artif. Intell., 18, 157-174.
•  Dufour,  É.,  Frencia,  J. P., and  Kane,  E. (2003). Development of a rapid method based on front-face fluorescence spectroscopy for the monitoring of fish freshness. Food Res. Int., 36, 415-423.
•  ElMasry,  G.,  Nagai,  H.,  Moria,  K.,  Nakazawa,  N.,  Tsuta,  M.,  Sugiyama,  J.,  Okazaki,  E., and  Nakauchi,  S. (2015). Freshness estimation of intact frozen fish using fluorescence spectroscopy and chemometrics of excitation–emission matrix. Talanta, 143, 145-156.
•  Harshman,  R. A. and  Lundy,  M. E. (1994). PARAFAC: Parallel factor analysis. Comput. Stat. Data Anal., 18, 39-72.
•  Hasegawa,  K.,  Endo,  Y., and  Fujimoto,  K. (1992). Oxidative deterioration in dried fish model systems assessed by solid sample fluorescence spectrophotometry. Journal of food science, 57, 1123-1126.
•  Karoui,  R. and  Blecker,  C. (2011). Fluorescence spectroscopy measurement for quality assessment of food systems—a review. Food Bioprocess Technol., 4, 364-386.
•  Karoui,  R.,  Kemps,  B.,  Bamelis,  F.,  De Ketelaere,  B.,  Merten,  K.,  Schoonheydt,  R.,  Decuypere,  E., and  De Baerdemaeker,  J. (2006a). Development of a rapid method based on front face fluorescence spectroscopy for the monitoring of egg freshness: 1—evolution of thick and thin egg albumens. Eur. Food Res. Technol., 223, 303-312.
•  Karoui,  R.,  Kemps,  B.,  Bamelis,  F.,  De Ketelaere,  B.,  Merten,  K.,  Schoonheydt,  R.,  Decuypere,  E., and  De Baerdemaeker,  J. (2006b). Development of a rapid method based on front-face fluorescence spectroscopy for the monitoring of egg freshness: 2—evolution of egg yolk. Eur. Food Res. Technol., 223, 180-188.
•  Murphy,  K. R.,  Stedmon,  C. A.,  Graeber,  D., and  Bro,  R. (2013). Fluorescence spectroscopy and multi-way techniques. PARAFAC. Anal. Methods, 5, 6557-6566.
•  Nakamura,  Y.,  Fujita,  K.,  Sugiyama,  J.,  Tsuta,  M.,  Shibata,  M.,  Yoshimura,  M.,  Kokawa,  M.,  Nabetani,  H., and  Araki,  T. (2012). Discrimination of the geographic origin of mangoes using fluorescence fingerprint. J. Jpn. Soc. Food Sci. Technol. (Nippon Shokuhin Kagaku Kogaku Kaishi), 59, 387-393 (in Japanese).
•  Nishino,  K.,  Nakamura,  K.,  Tsuta,  M.,  Yoshimura,  M.,  Sugiyama,  J., and  Nakauchi,  S. (2013). Optimization of excitation–emission band-pass filter for visualization of viable bacteria distribution on the surface of pork meat. Opt. express, 21, 12579-12591.
•  Olsen,  E.,  Veberg,  A.,  Vogt,  G.,  Tomic,  O.,  Kirkhus,  B.,  Ekeberg,  D., and  Nilsson,  A. (2006). Analysis of early lipid oxidation in salmon pâté with cod liver oil and antioxidants. J. Food Sci., 71, S284-S292.
•  Ram,  M. S.,  Seitz,  L. M., and  Dowell,  F. E. (2004). Natural fluorescence of red and white wheat kernels. Cereal Chem., 81, 244-248.
•  Rinnan,  Å. and  Andersen,  C. M. (2005). Handling of firstorder Rayleigh scatter in PARAFAC modelling of fluorescence excitation–emission data. Chemom. Intell. Lab. Syst., 76, 91-99.
•  Saito,  T.,  Arai,  K., and  Matsuyoshi.  M. (1959). A new method for estimating the freshness of fish. B. Jpn. Soc. Sci. Fish, 24, 749-750.
•  Savitzky,  A. and  Golay,  M. J. (1964). Smoothing and differentiation of data by simplified least squares procedures. Anal. Chem., 36, 1627-1639.
•  Sorrell,  M. J.,  Tribble,  J.,  Reinisch,  L.,  Werkhaven,  J. A., and  Ossoff,  R. H. (1994). Bacteria identification of otitis media with fluorescence spectroscopy. Lasers Surg. Med., 14, 155-163.
•  Sterenborg,  H. J. C. M.,  Motamedi,  M.,  Wagner,  R. F.,  Duvic,  M.,  Thomsen,  S., and  Jacques,  S. L. (1994). In vivo fluorescence spectroscopy and imaging of human skin tumours. Lasers Med. Sci., 9, 191-201.
•  Trivittayasil,  V.,  Tsuta,  M.,  Kasai,  S.,  Matsuo,  Y.,  Sekiyama,  Y.,  Shoji,  T.,  Aiyama,  R.,  Kokawa,  M. and  Sugiyama,  J. (2018). Classification of 1-methylcyclopropene treated apples by fluorescence fingerprint using partial least squares discriminant analysis with stepwise selectivity ratio variable selection method. Chemom. Intell. Lab. Syst., 175, 30-36.
•  Warner,  I. M.,  Christian,  G. D.,  Davidson,  E. R., and  Callis,  J. B. (1977). Analysis of multicomponent fluorescence data. Anal. Chem., 49, 564-573.
•  Werkhaven,  J. A.,  Reinisch,  L.,  Sorrell,  M.,  Tribble,  J., and  Ossoff,  R. H. (1994). Noninvasive optical diagnosis of bacteria causing otitis media. The Laryngoscope, 104, 264-268.
•  Wolf,  G.,  Almeida,  J. S.,  Pinheiro,  C.,  Correia,  V.,  Rodrigues,  C.,  Reis,  M. A., and  Crespo,  J. G. (2001). Two-dimensional fluorometry coupled with artificial neural networks: A novel method for on-line monitoring of complex biological processes. Biotechnol. Bioeng., 72, 297-306.