A.I. Ropodi, D.E. Pavlidis, D. Loukas, P. Tsakanikas, E.Z. Panagou and G.-J.E. NYCHAS E-mail: gjn@aua.gr
..an alternative approach is needed within the PAT concept
This work aims to investigate the potential of combining data from multispectral images and FTIR measurements in the context of PAT in the meat sector in an attempt to estimate the microbial population on meat directly from spectral data.
Product: Minced beef (75-80 g portions). Packaging: aerobic and modified atmosphere packaging (MAP, 80%O 2 /20%CO 2 ). Storage temperature: 5 C and 10 C. 51 samples (3 control + 12 for each case of temperature & packaging). All samples were analyzed microbiologically for the determination of Total Viable Counts (TVCs). In parallel, spectral data were acquired through multispectral imaging and FTIR. TVCs range during storage: 5.68-8.75 log CFU/gr.
Multispectral images were acquired using VideometerLab, Wavelengths ranging from 405-970nm (visible & NIR). Wavelengths (nm) 405 435 450 470 505 525 570 590 630 645 660 700 850 870 890 910 940 970
Reflectance Oxymyoglobin Myoglobin Metmyoglobin Water Fat Wavelength (nm)
Image segmentation: an in-house MATLAB script * for automated image segmentation based on Gaussian Mixture Models (GMM)** was used in order to remove background and fat. For each wavelength the mean and standard deviation of the pixels intensity value was acquired for further processing, resulting in 36 variables. *Ropodi, A.I., Tsakanikas, P., Loukas, D., Panagou, E.Z. and Nychas, G. J.E. (2014) Multispectral image analysis for the assessment of pork minced meat quality and mapping of microbial contamination. EUROPT(R)ODE XII Conference on Optical and Chemical Sensors, Athens - Greece, 13-16th April **McLachlan, G.J. and Peel, D., (2000). Finite Mixture Models, Wiley
JASCO 6200 FTIR spectrometer for acquisition of spectra: Standard Normal Variate (SNV) preprocessing was applied. As the spectra include a large number of wavenumbers/variables that can be noisy and collinear, downsampling was performed by taking the average spectrum of every 4 wavenumbers. The resulting variables were reduced to 830.
The possibility of combining data from both sensors (data fusion) was explored. The methods were based mainly on Partial Least Squares (PLS), by applying regression or by using the PLS components as an intermediate step. For calibration, the Mean Square Error (MSE) for Leave-One- Out Cross-Validation (LOOCV) was used. A multi-objective genetic algorithm coupled with PLS (multiobjective GA-PLS)* was utilized for variable selection, minimizing the number of variables and MSE of CV. *Loukas, D., Ropodi, A., & Nychas, G.-J. (2014). Regression modeling for spectral data sets : A multi-objective genetic approach. In 3 rd International Symposium & 25 th National Conference on Operational Research Proceedings (pp. 140 147). ** NSGA-II algorithm (available in MATLAB)
As raw data are extremely complex and noisy, the fusion methodologies were applied: At data level (after data pre-processing). At feature level
1. 2. 3. VM data FTIR data VM data FTIR data VM data FTIR data fusion fusion PLS variable selection GA-PLS VM variables after selection with GA-PLS FTIR variables after selection with GA-PLS fusion Final estimation Final estimation PLS Final decision
4. 5. VM data FTIR data VM data FTIR data Selected PLS components Selected PLS components Selected PLS components Selected PLS components fusion PLS variable selection fusion GA-PLS Final estimation Final estimation
predicted predicted TVC TVC 9 8.5 8 7.5 7 6.5 6 5.5 5.5 6 6.5 7 7.5 8 8.5 9 observed 5.5 5.5 6 6.5 7 7.5 8 8.5 VM only FTIR only MSE CV 0.4731 0.2477 Bias 1.0021 1.0003 Accuracy 1.0535 1.0164 RMSE 0.4623 0.1488 SEC(%) 6.3914 2.0575 8.5 8 7.5 7 6.5 6 observed
predicted predicted predicted 9 8.5 8 7.5 7 VM only FTIR only Case 1 (PLS) Case 2 (GAPLS) Case 3 (GA-PLS& PLS) MSE CV 0.4731 0.2477 0.2513 0.1671 0.1927 Bias 1.0021 1.0003 1.0009 1.0008 1.0004 Accuracy 1.0535 1.0164 1.0330 1.0301 1.0190 RMSE 0.4623 0.1488 0.3024 0.2829 0.1722 SEC(%) 6.3914 2.0575 4.1805 3.9105 2.3811 TVC 9 8.5 8 7.5 7 TVC 9 8.5 8 7.5 7 TVC 6.5 6.5 6.5 6 6 6 5.5 5.5 6 6.5 7 7.5 8 8.5 9 observed 5.5 5.5 6 6.5 7 7.5 8 8.5 9 observed 5.5 5.5 6 6.5 7 7.5 8 8.5 9 observed
predicted predicted VM only FTIR only Case 4 (PLS comp. & PLS) Case 5 (PLS comp. & GAPLS) MSE CV 0.4731 0.2477 0.0748 0.0494 Bias 1.0021 1.0003 1.0002 1.0003 Accuracy 1.0535 1.0164 1.0162 1.0178 RMSE 0.4623 0.1488 0.1401 0.1654 SEC (%) 6.3914 2.0575 1.9371 2.2863 8.5 8 7.5 7 6.5 TVC 8.5 8 7.5 7 6.5 TVC 6 6 5.5 5.5 6 6.5 7 7.5 8 8.5 observed 5.5 5.5 6 6.5 7 7.5 8 8.5 observed
Two very different types of data were combined with VideometerLab data being less accurate. Sometimes the combination of data yielded superior results compared to data from only one sensor. Best cases were 3 (GA-PLS & PLS), 4 (PLS comp. & PLS) as they were the only ones where all samples were within the ±0.5 log cycle range.
Apply other data analysis techniques (except PLS) to improve fusion. Validate model with new experimental data. Incorporate other sensors. Build an intensive database that will incorporate as much variability of the meat as possible.
www.imeatsense.gr This work has been supported by the project Intelligent multi-sensor system for meat analysis - imeatsense_550 co financed by the European Union (European Social Fund ESF) and Greek national funds through the Operational Program "Education and Lifelong Learning" of the National Strategic Reference Framework (NSRF) Research Funding Program: ARISTEIA I.