Image analysis approaches to colorimetric and multispectral detection of heavy metal or bacterial contamination with printed paper-based test devices

Nowadays, safety in food has become critical. The main two types of threats related to food safety are foodborne pathogens and heavy metals. One of the most common foodborne pathogens that can be found in our daily food is Escherichia coli O157:H7 (\emph{E. coli} O157:H7). Human infections with \emph{E. coli} O157:H7 poses severe disease in our bodies and even death. Heavy metals, Mercury (Hg), Arsenic (As), Copper (Cu), and so on can be enriched in living tissue through food chains and have proven to be harmful to human health at low concentrations. So, the detection of the contaminants in daily food and drinking water is crucial for global public health. To date, the widely used pathogen detection methods and heavy metal detection methods are expensive, involve laborious procedures, and cannot support on-site detection. In this dissertation, two different novel printing platforms, and accompanying rapid, microfluidic architecture platforms are proposed for the detection of \emph{E. coli} O157:H7, and As and Hg, respectively. These devices can be fully integrated with a mobile phone and an image analysis pipeline to capture and analyze the sensor images on-site.

The first system that we develop consists of inkjet printed lateral diffusion strips for the capture of \emph{E. coli} O157:H7. This effort includes the characterization of the consistent size droplets used to develop the print masks, and functional printing of the test and control lines with the corresponding DNA solutions. Then we propose an image analysis method to read the responses of the test strips to the presence of targets, and quantitatively correlate the color intensity of responses to the target concentrations. Finally, the response variation of the optical properties of the test lines detecting the target at the same concentration is investigated. To reduce the sample-to-sample variation in response to \emph{E. coli} O157:H7, and in an effort to reduce material cost and printing time, we optimize the numbers of print layers and explore various well-used image segmentation methods to detect the response in the test lines of test strips. The usefulness of these segmentation methods is evaluated by comparing the response variance of the corresponding segmentation results.

To move towards multiplexed heavy metals detection of As and Hg, we design a second patterned microfluidic paper-based device, capable of instrument-free, portable, and multiplexed sensing detection via aptamer recognition. We explore three different printing technologies for fabricating these devices, and choose screen printing with UV cured inks for further development. We then conduct empirical studies to optimize the device geometry. We propose three image analysis methods to obtain a higher prediction accuracy with our developed paper-based devices for detecting and measuring heavy metal contaminants in food or liquids. (1) We use $\Delta$E from a white background as our baseline method to correlate the optical properties with the different concentrations of the target, and optimally quantize these responses into five groups to evaluate the prediction accuracy. (2) We propose a CNN classifier and explore two kinds of data augmentation techniques to compare their effectiveness for the classification task. (3) We consider the use of the spectral reflectance of the sensor pad, then develop two different machine learning approaches: k-nearest-neighbor with sequential forward feature selection to determine the best number of features, and random forest with principal component analysis for feature reduction for classifying the level of contamination by As$^{3+}$ into one of five categories. The accuracy of these three models is compared by implementing them with the same training and test datasets.