Volume 55, Number 8, 2001 APPLIED SPECTROSCOPY 1025 0003-7028 / 01 / 5508-1025$2.00 / 0 q 2001 Society for Applied Spectroscopy Detection of Parasites in Cod Fillets by Using SIMCA Classication in Multispectral Images in the Visible and NIR Region JENS PETTER WOLD,* FRANK WESTAD, and KARSTEN HEIA MATFORSK Norwegian Food Research Institute, A Ê s, Norway (J.P.W., F.W.); and Norwegian Fisheries and Aquaculture Research Institute Ltd., Tromsø, Norway (K.H.) The presence of parasitic nematodes in llets of commercially im- portant sh species has been a serious quality problem for the sh- ing industry for several decades. Various approaches have been tried to develop an efcient method to detect the parasites, but so far the only reasonable solution is manual inspection and trimming of each sh llet on a candling table. In this study we have inves- tigated how multispectral imaging in combination with SIMCA classication can be used for automatic detection of parasites. The results indicate that the spectral characteristics of nematodes differ sufciently from those of sh esh to allow one to obtain fairly good classications. The method is able to detect parasites at depths down to about 6 mm into the sh muscle. The method shows promising results, but further studies are required to verify feasibility for the sh industry. Index Headings: Fish quality; Cod worms; Nematodes; Automatic detection; Multispectral imaging; Principal component analysis; PCA; SIMCA classication. INTRODUCTION The presence of parasitic nematodes (cod worms, seal worms, whale worms) in llets of various sh species has been a serious quality problem for the shing indus- try for several decades. The sh (e.g., cod, herring, coal- sh) serve as intermediate hosts for the parasites. In the sh, the worms can migrate from the intestines into the sh muscle, where they stay until the sh is eaten by a primary host (seal or whale). Human consumption of dead nematodes is harmless, so in countries where it is common to either boil, freeze, fry, or heavily salt the sh before consumption, it is not a public health problem. For the consumers, however, the presence of nematodes is unacceptable. To fulll market requirements and avoid complaints, the sh industry must be able to deliver par- asite-free products. Several approaches have been tried in the effort to develop an efcient method to detect the parasites, but so far, the only reasonable solution has been manual inspection and trimming of each llet on a can- dling table. 1 This process is very laborious and costly, and even with optimal conditions, not more than about 75% of the parasites can be expected to be found. 2 A technique for automatic detection of parasites in sh l- lets is therefore much needed. Several principles of detection have been tested [visi- ble light, IR light, uorescence, X-ray, ultrasound, com- puter tomography (CT) and magnetic resonance imag- ing], 3–5 but only CT and ultrasound seem to be able to detect deeply embedded parasites. CT is still too expen- Received 8 February 2001; accepted 24 April 2001. * Author to whom correspondence should be sent. sive for the industry. Ultrasound seems unpractical be- cause the method requires close physical contact between llet and transducers, and the method is very sensitive to air bubbles in/on the llets. Focus of the research has therefore been on optical methods. A problem with the use of trans-illumination by visible light is the very prominent light scattering in sh muscle, which makes it difcult to detect objects inside the llets. Visible light gives contrast between sh muscle and parasites, but by manual inspection, the worms cannot be detected deeper than about 6 mm. 6 Petursson 6 studied the optical prop- erties of sh muscle and reported that absorption spectra from parasites and white and dark sh muscle, as well as blood-stained muscle, showed differences most markedly in the 400–600 nm range. He also showed that scattering decreases with increasing wavelength. He concluded that the near-infrared (NIR) area from 800 to 1200 nm could give better results for deeply embedded parasites. In light of these ndings, it is reasonable to assume that detection of parasites can be based on spectral characteristics. Multispectral imaging has become a commonly used technique within elds spanning all the way from mi- croscopy and medical imaging to satellite remote sensing. It is a powerful basis for segmentation and classication, and it can be used to visualize the chemical composition of materials. 7 Applications have also been suggested for the agriculture and food industries. On-line detection of fecal contamination on poultry carcasses, 8 identication of food products such as maize, pea, soy, and wheat, 9 detection of defects on apples, 10 and mapping of lipid oxidation in chicken meat 11 are some reported applica- tions. Human beings are excellent at detecting parasites. With the combination of color and morphological features with a priori experience, it is rather easy to distinguish visible worms from the sh esh. A computer-based detection system using only morphological image features will probably have limited performance, because the parasites can appear in any shape and are often very similar to features in the sh llet. Spectral properties are indepen- dent of the parasite’s shape. The objective of the work presented in this paper was therefore to investigate whether parasites could be distinguished in cod llets purely on the basis of spectral characteristics. To do this, we collected multispectral images of parasite-infested cod llets in the visible and near-infrared. Modeling and clas- sication were performed by use of multivariate image analysis, including SIMCA classication (soft indepen- dent modeling of class analogies). The paper describes