Application of an artificial neural network in the enumeration of yeasts and bacteria adhering to solid substrata

Abstract

Artificial neural networks (ANNs) combined with automated image processing are being used in a growing number of applications, ranging from car license plate identification to speech recognition. ANN analysis is capable of handling complicated images that cannot be dealt with using conventional image filters, such as images originating from yeasts and bacteria adhering on turbid solids (silicone rubber or dental enamel) or metals. In this paper, a feed-forward ANN processing greyvalues from a 9×9 pixel microscope subimage, is presented for the image enhancement and subsequent enumeration of microorganisms adhering to solid substrata. The ANN developed is a simple software routine that can be implemented as a digital image filter in existing software and assists in locating the positions of adhering microorganisms. In order to evaluate the performance of the ANN, adhesion experiments were carried out in a parallel plate flow chamber with Candida albicans, Enterococcus faecalis, Streptococcus thermophilus, Pseudomonas aeruginosa, Staphylococcus epidermidis, and polystyrene latex particles to different solid substrata under ideal and subideal illumination and focus of the microscope setup. The performance of the ANN-based enumerations was compared with the performance of nonneural, conventional image-processing methods and visual enumeration. For high-contrast images, nearly all of the adhering organisms can be detected correctly after conventional image filtering. However, in the lower quality images which were employed in this study, for metal or silicone rubber substrata, image enhancement by the ANN yielded enumeration of the adhering bacteria with an accuracy of 93%–98%, while using conventional image filters, an accuracy of 86–96% was achieved. Adhering yeasts could be enumerated after conventional image filtering with an accuracy of 92%, but after ANN enhancement, 98% of the yeast cells adhering on silicone rubber substrata were enumerated correctly. It can be concluded that for low quality and complicated images, ANNs perform better than conventional filters, while as additional advantages, the training of an ANN requires significantly less expert knowledge than optimal setting of conventional image filters, and an ANN can be used to distinguish between strains in adhesion experiments involving more than one strain.