Analyzing Uncertainty in Microplastic Detection: A Comprehensive CT Scan and Neural Network Approach

Microplastics (MPs), smaller than 5 mm and often invisible to the naked eye, represent a significant threat to marine ecosystems and human health. Traditional methods for detecting MPs often lack efficiency and may compromise sample integrity. In this study, a novel non-destructive methodology for detecting MPs in fish samples is proposed, utilizing computed tomography (CT) scanning and neural networks (NN). Leveraging recent advancements in uncertainty analysis, the reliability of the innovative approach and its application to fish sample analysis is assessed. Sources of uncertainty originating from both tomographic imaging and NN processing are identified, addressing issues such as false positives/negatives and uncertainties in measured parameters such as particle position, size, and material. Through post-processing techniques and calibration procedures, uncertainties are mitigated, providing insights into epistemic and stochastic uncertainties inherent in the methodology. Our findings demonstrate that the calibration system used in conjunction with tomographic imaging shows good linearity, with a standard deviation of 0.029 kg/m3. Despite inherent noise in the images causing variability in gray levels, the range of variability for the plastics of interest was narrower than the uncertainty of the calibration curve. This indicates that while the calibration was accurate, the inherent noise still posed challenges for precise material classification. This comprehensive analysis enhances the reliability and accuracy of MPs detection in fish samples, contributing to a deeper understanding of uncertainty in CT-based analyses and NN.