Deep learning-powered quantitative test-strip platform enables robust detection of analytes in crudely extracted samples via a cloud-based smartphone application

Point-of-care testing (POCT) faces challenges in robustly detecting trace analytes because of its limited ability to distinguish weak target signals from background noise, resulting in a risk of false-negative results. We present a deep learning (DL)-driven POCT assay based on commercial strips that allows for the robust detection of crudely extracted samples, and thus more reliable and quantitative detection compared with conventional intensity-based analysis. The assay leverages a target-responsive DNA hydrogel functionalized with recognition molecules and offers a general principle for the detection of diverse analytes. Rather than conventional intensity-based image analysis, the assay employs a DL tool as the signal readout to mitigate test variability and enhance subtle signals. Furthermore, a cloud-based smartphone application streamlines the user experience to provide intuitive and reliable detection results within seconds. We show that the DL-powered approach effectively distinguishes low-concentration positive signals and background, achieving over 90.0% accuracy, and demonstrating improved performance compared with traditional grayscale analysis. It significantly enhances sensitivity by 2- to 5-fold and expands the detection range by an order of magnitude for small-molecule mycotoxins. The platform enables the rapid and accurate quantification of mycotoxin in diverse contaminated food samples, with an average accuracy of 92.4%. This approach demonstrates strong agreement with standard laboratory methods in complex sample analysis and an improvement in POCT applicability compared with conventional readout methods.