Classification of Foods and Plastics using FD-FLIM and Neural Networks

Plastics and foods can be differentiated by their material characteristic fluorescence properties, especially their fluorescence lifetimes. An areal measurement of fluorescent lifetimes of these materials can be done using Frequency-Domain Fluorescence Lifetime Imaging Microscopy (FD-FLIM). Up until now, most plastic detection is done using NIR or X-ray, while most applications of FD-FLIM are in biomedicalfields. The application of FD-FLIM in a food safety setting presents a promising approach to the detection of plastic contaminants. A Multilayer Perceptron (MLP) based neural network is developed to reliably identify the presence of plastic in a food/plastic sample via FD-FLIM. Features like the mean, median, standard deviation, variance, range, and interquartile range are calculated from the intensity image, the phase shift and modulation index along with the according phase- and modulation-dependent fluorescence lifetimes from the FD-FLIM data. For training, test and validation, a total of 3520 FD-FLIM measurements have been taken at 445nm excitation of sixteen samples with the labels food and plastic. To rank the performance of the 3888 trained networks, Fl-score, accuracy, precision, and recall are used as metrics. The best performing network reaches a Fl-score of 98.86% proving that a differentiation of foods and plastics using a MLP classification based on FD- FLIM data is possible with a low error rate.