Integration Of Spectral-spatial Information For The Identification Of Meat And Bone Meal Using NIR Microscopic Imaging

Meat and bone meal (MBM) is one of the main prion-contaminated animal feed ingredients responsible for BSE, generally called "mad cow disease". As many countries have legislated the banning of MBM as a feedstuff or feed ingredient to prevent the transmission of this disease among cattle and other livestock, techniques to detect MBM are required. By combining near-infrared (NIR) spectroscopy and microscopy, NIR microscopic imaging records a spectrum per site (pixel) of sample surface and finally forms an image of the sample. This new advanced analytical technique is a rapid, non-destructive, non-polluted, high-accuracy and micro-visualization method in characterizing complex mixtures. This method will be enhanced for the identification of MBM by extracting spectral information combined with spatial information from NIR spectroscopic images. This study investigated the appropriate approaches for integrating spatial with spectral information from NIR microscopic image. The research work has important scientific and practical value in enriching rapid detection method of MBM and preventing the occurrence of BSE disease.1. Samples of MBM, dairy concentrate supplement and fish meal (FM) were collected, sedimented and arranged on polytetrarruoroethene (PTFE) background plate for NIR microscopic imaging. Both principal component analysis (PCA) and K-means clustering analysis were used to extract and present relevant information from NIR microscopic images. The results showed that MBM could be distinguished from dairy concentrate supplement/FM by the scores from principal component analysis, and samples were subdivided by using K-means clustering based on the PCA analysis. It is demonstrated that NIR microscopic imaging approach is one of most promising methods for the detection of MBM.2. To test the performance of NIR microscopic imaging to species identify MBM, a rapid method to construct NIR microscopic standard spectra database is demonstrated. For the study, bone fragments of three different species MBM (porcine origin, avian origin and ruminant origin) were analyzed on NIR microscopic imaging system, and both VIS and NIR images were acquired at the same size. To extract and mark the position of every single bone fragment in visible image, a graphical user-friendly interface, based on marker-controlled watershed segmentation method, written in Matlab for extracting and marking bone fragments has been developed in this paper. By the position information in VIS image, NIR microscopic image was decompounded to spectrum of each bone fragment. The determination model is then constructed by two methods (PLSDA and SVMDA) with different spectral preprocessing. The sensitivity of the best discrimination model is0.896,0.949and0.918, respective for three species. The specificity is0.963,0.969and0.950.3. This study introduces an innovative approach to analyzing NIR microscopic images: an Markov random field-based approach has been developed using the ICM (Iterative conditional mode) algorithm, integrating initial labeling derived results from SVMDA and observation data derived from the results of PCA. The results showed that MBM covered by feed can be successfully recognized with an overall accuracy of86.59%and a Kappa coefficient of0.68. MBM covered by FM can be successfully recognized with an overall accuracy of86.22%and a Kappa coefficient of0.66.Compared with conventional methods, the MRF-based approach is capable of extracting spectral information combined with spatial information from NIR microscopic images. This new approach enhances the identification of MBM using NIR microscopic imaging.4. Feed containing between0%-100%(w/w) MBM are prepared to test the quantification ability of NIR microscopic imaging. Three different strategies are demonstrated to construct a quantification model, including PLSR, SVMDA and MRF-based approach. The results showed that the determination coefficients (R2) of three quantification models for the low level content samples (0%MBM-10%MBM) is0.52,0.56and0.62with an average absolute error (AAE)0.89%,0.97%and0.60%, respectively. The determination coefficients (R2) of three quantification models for all samples is0.99,0.96and0.98with an AAE0.46%,0.59%and0.42%, respectively.