|Hsieh, Ching-Lu - ISL VISITING SCIENTIST|
|Dey, Bhabani - USDA, FSIS, WASHINGTON DC|
Submitted to: Transactions of the ASAE
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: December 11, 2001
Publication Date: March 1, 2002
Citation: HSIEH, C., CHEN, Y.R., DEY, B.P., CHAN, D.E. SEPARATING SEPTICEMIC AND NORMAL CHICKEN LIVERS BY VISIBLE/NEAR INFRARED SPECTROSCOPY AND BACK-PROPAGATION NEURAL NETWORKS. TRANSACTIONS OF THE AMERICAN SOCIETY OF AGRICULTURAL ENGINEERS. 45(2):459-469. 2002. Interpretive Summary: Previously, we have developed a Vis/NIR spectroscopy technique for on-line classification of wholesome and unwholesome chicken carcasses. Testing of a pilot-scale Vis/NIR spectroscopy system at a poultry processing plant showed that the system can achieve 95% accuracy in separating wholesome and unwholesome carcasses. In order to further improve the effectiveness of the Vis/NIR technique, sensing of visceral organs for separating wholesome and unwholesome carcasses was attempted. Since, under the current FSIS's HACCP-based inspection model project, any carcasses showing septicemic symptoms should not be allowed to reach marketplace, the object of this research was to study the feasibility of using Vis/NIR techniques for the separation of septicemic livers from the normal livers. Back-propagation neural networks were used to develop classification models. Different spectral data preprocessing methods were compared. The laboratory study showed that the best model achieved a classification accuracy of 98% for normal livers and 94% for septicemia livers. This information is very important to the FSIS and to any scientists and engineers who are developing sensors to detect animal diseases.
Technical Abstract: Visible/near-infrared spectra of three hundred chicken livers were analyzed to explore the feasibility of using spectroscopy to separate septicemic livers from normal livers. Three strategies involving offset, second difference, and functional link methods were applied to preprocess spectra, while principal component analysis (PCA) was utilized to reduce the input data dimensions. PCA scores were fed into a feed forward back-propagation neural network for classification. The results showed no obvious difference in classification accuracy between offset and non-offset data when no other preprocessing method was applied. The full 400-2498 nm wavelength region produced better results than the 400-700 nm, 400-1098 nm, and 1102-2498 nm sub-regions. In general, the classification accuracy was improved by increasing the number of scores of input data, but too many scores also diminished performance. The functional link test showed that using functional link spectra selected at every third point with 60 scores achieved the same classification accuracy as that obtained when using all the data points with 90 scores. The best classification model used offset correction followed by second difference (g=31) and 60 scores. It achieved a classification accuracy of 98% for normal and 94% for septicemia livers.