Abstract There is a significant global health need to translate more in vitro diagnostic tests from clinical laboratories to field-based applications, including point-of-care and self-administered test formats. These applications typically require smaller sample sizes, limit sample processing and measurement capabilities, and introduce greater handling variability. Error tolerance is one of the most critical factors for successful field-based assay design. Here, we examine machine-learning (ML) strategies to enhance the error tolerance of image-based nanoparticle immunoassays. Random dispersions of nanoparticles were imaged in microliter sample volumes, and images were processed to determine analyte concentrations based on nanoparticle appearance. Assay performance was characterized using two common blood analytes: C-reactive protein and anti-SARS-CoV-2 IgG. We compare the results from conventional image analysis, a hybrid ML-conventional approach based on pixel segmentation, and end-to-end image regression using a targeted regularization strategy. Using serum samples from SARS-CoV-2 positive individuals, the segmentation-based approach enabled binary classification with 96% specificity and 90% sensitivity, matching seroconversion rates. The end-to-end regression model achieved superior quantitative performance (5.2 ng/mL), approaching ELISA-level detection range (0.01–10 ng/mL, depending on capture antibody affinity) in a single 30 min workflow without sample preprocessing. The limit of detection for digital molecular assays is not fixed, and we perform a theoretical analysis showing how adjusting particle counts and polydispersity can achieve arbitrary sensitivity down to the Poisson limit. Training images for the full image regression approach required only a single label—the analyte concentration—eliminating labor-intensive pixel-level labeling. Ultimately, the image-based readouts significantly improved dynamic range, sensitivity, and reproducibility over conventional readouts.
McAffee et al. (Mon,) studied this question.