Towards Portable and Accurate Quantification of Lateral Flow Assays Using Smartphone Imaging and Deep Learning

Lateral flow assays (LFAs) integrated with smartphone imaging hold potential for next-generation portable diagnostic platforms. Conventionally, LFAs such as pregnancy tests and COVID-19 rapid antigen kits provide only qualitative YES/NO answers and are not designed for quantitative analysis. Although recent advances have attempted to enable quantification in LFAs, major challenges remain, including variability in image acquisition, reliable region-of-interest (ROI) detection, limited availability of annotated datasets, and achieving accurate as well as precise quantification. To address these limitations, we present a generic algorithm designed to be compatible across diverse LFA platforms. The software, driven by ResNet-50 convolutional neural networks (CNNs), scans and extracts relevant image descriptors for diagnosis, eliminating the need for manual ROI uploading and marking. Various commercially available diagnostic kits were tested, with images captured both in a controlled in-house imaging box and under conventional environmental conditions across multiple smartphones. The resulting dataset was used for training the CNN model ResNet-50, which demonstrated the highest accuracy (76.2%), MSE (0.00291), and R² (0.9398), outperforming others, which showed reduced robustness under image augmentation and variable lighting conditions. Additionally, a comparative evaluation of different CNN models confirms the superior adaptability of ResNet-50 for real-world imaging conditions. As a proof of concept, the method was further implemented for smartphone-based quantification of vitamin D3 levels using an in-house developed assay. Collectively, these findings lay the groundwork for accurate, quantitative, and portable diagnostics, useful in low-resource settings.