Lateral Flow Test Interpretation with Residual Networks

Abstract

Lateral flow tests (LFTs) are a cost-effective, quick, and frequently used testing method in many domains such as food safety and environmental and clinical applications. However, a major challenge is accurate interpretation of LFT results. Often, if a low level of target substance is present in the input liquid, the test line indicator may appear faint, causing a test interpreter to read the test result as a false negative. Therefore, to address this problem, we propose a deep-learning-based method to interpret images of LFT results. Our model is based on ResNet-101, a state-of-the-art image classification model that uses residual networks, or skip-connections between layers to improve learning on the training dataset. We further improve our model by using data augmentation to generate additional and more difficult images of LFTs for the model to learn from, thereby improving its performance and reducing overfitting to the training dataset. Our approach is also trained and tested on a dataset of SARS-CoV-2 LFT images, containing both positive and negative results. We compare our ResNet approach to a baseline convolutional neural network model. Our results show the ResNet model achieves a higher specificity and sensitivity than the baseline model to interpret LFT results.