Rapid antigen tests became one of the defining tools of the COVID-19 pandemic, but their readout has always rested on a human eyeball: a faint second line, judged positive or negative by whoever happens to be holding the card. A team of researchers in the United States has now built a computer vision algorithm that reads the Abbott BinaxNOW COVID-19 Ag card objectively, estimates the viral load behind the result, and predicts roughly when a user who tests positive will test negative on the same device. The work, published in BMC Digital Health, packages the analysis into a smartphone application intended to make home testing more informative than a simple yes-or-no answer.
The motivation stems from well-known weaknesses of visually read lateral flow assays. These tests detect viral antigens in minutes and at low cost, which made them the default screening tool for households, schools and events. Compared with real-time RT-PCR, however, they are less sensitive and less specific, they return only a qualitative result, and they depend on the user interpreting the lines correctly, particularly when the test line is faint. Public health guidance during the pandemic increasingly asked laypeople to act on their own interpretations, raising the stakes of those judgments. The researchers set out to reduce interpretation error, to correlate the test line with PCR cycle threshold values as a semi-quantitative proxy for viral load, and to forecast the date of a negative result to guide decisions about ending isolation precautions.
The image processing pipeline begins with an ordinary photograph of the test card. The algorithm, written in Python with the openCV library, exploits landmarks printed in magenta ink on the BinaxNOW card. A color difference image, calculated as red plus blue minus green, separates the magenta header from the rest of the photo; Otsu binarization then yields a bounding rectangle of the header, which is used to crop and rotate the card into a portrait orientation. Three magenta lines bracketing the test strip are extracted the same way, and the strip itself is cropped from the region between and above them. Plotting the intensity of the color difference image across the strip reveals peaks where gold nanoparticle-tagged antibody has accumulated, marking the control and sample regions of interest that the model must analyze.
To quantify those lines under varying lighting, the team developed a measure they call accumulated line intensity, or ALI. The mean grayscale intensity is computed across each pixel row of the test strip. A baseline intensity, representing the strip without bound indicator, is interpolated by linearly fitting the mean of the top and bottom twenty percent of each region of interest. The difference between this estimated baseline and the observed intensity is then summed to produce the ALI for the control and sample lines, and the ratio of sample to control ALI serves as a normalized metric for each test. This normalization is what allows images taken freehand in ambient light to be compared with images taken under controlled conditions.
The relationship between the ALI ratio and laboratory measurements was established with spiked samples. Remnant clinical SARS-CoV-2 samples were pooled and diluted in a twofold series with negative nasal wash; N2 cycle threshold values were determined with the CDC assay and antigen concentrations were measured on a Quanterix HD-X platform. Samples were spiked onto BinaxNOW cards and imaged with an iPhone 13 Pro placed on top of a commercial lightbox with its internal light at full brightness, using automatic exposure and a two-second shutter timer. Two regression models were fitted on eighteen laboratory samples: an exponential model with cycle threshold as the outcome and a sigmoidal dose-response curve with antigen concentration in pg/mL as the outcome, in both cases using the ALI ratio as the lone predictor.
Validation on a separate set of thirty blinded laboratory samples showed that the cycle threshold model performed better than the antigen model. Predicted CT values tracked actual laboratory values with a root mean square error of 1.44 CT, while predicted antigen concentrations deviated by an RMSE of about 21,634.87 pg/mL. Six samples with an associated CT of 0.0 were excluded from the RMSE calculations. The authors describe the CT and antigen predictions as correlating mildly well with the actual laboratory values, and they conclude that the CT model is the better predictor for this dataset, a caveat worth keeping in mind when considering the antigen quantification claims.
The more ambitious component is the time-to-negative model, built from a clinical substudy nested within the larger RADx infrastructure. Participants had to be enrolled in the parent study, present symptoms within zero to three days of enrollment, and test positive on a BinaxNOW device at enrollment; they were asked to return three to five days later. The substudy collected 118 baseline images from 59 subjects and 64 follow-up images from 32 subjects, with 16 negative and 16 positive results at follow-up. Average baseline CT was 22.3, ranging from 15.2 to 38.7; average follow-up CT was 30.7, ranging from 22.0 to 42.1; and participants averaged 1.9 days post symptom onset at baseline, with a range of zero to four days. Each test was photographed both inside the lightbox and freehand outside it.
A logistic mixed model predicted the BinaxNOW result from days since baseline and the baseline ALI ratio, with a random effect for each subject, fitted on the outside-the-lightbox images to simulate real-world use. Internal validation with Harrell’s bias correction produced an overall accuracy of 90 percent in predicting a patient’s status at follow-up, with a sensitivity of 96 percent and a specificity of 61 percent. The abstract reports somewhat higher figures, a sensitivity of 99 percent, specificity of 73 percent and accuracy of 94 percent, for predicting result status at a follow-up visit at least five days post symptom onset. The bias correction slightly decreased specificity, sensitivity and predictive accuracy, which the authors note may signal model overfitting, a real concern given that only about 90 data points were used. The outside-lightbox results also outperformed the controlled-lightbox versions, which the team reads as evidence of robustness.
The authors are candid about the limitations. Real-world users differ widely in technical proficiency, and the likelihood that laypeople can take images of similar quality to those captured by trained clinicians is, by their own admission, low; no home-use testing has yet been done. Lateral flow tests from other manufacturers have designs that may not match the magenta landmarks the vision model relies on, although the team believes the algorithms can be adapted with modest effort. The comparatively low specificity means some users will be told they are still positive and quarantine longer than necessary. The researchers deliberately prioritized sensitivity over specificity, reasoning that incorrectly clearing someone to rejoin others is the more dangerous error, and they call a large-scale, home-based study critical before any rollout.
Within those constraints, the technology points toward a different kind of home test. Because the algorithm quantifies the antigen present rather than merely detecting it, the authors argue it can serve as a proxy for viral load and disease course, information that visual reading cannot provide. The time-to-negative predictions are extrapolated day by day up to fourteen days from baseline, with 80 percent prediction intervals displayed, and a likelihood under 50 percent on a given day is read as a negative result on that day. The analysis was embedded in a native Android app written in Java, and the authors suggest that because lateral flow readouts are visually similar across diseases, the approach could extend beyond COVID-19. Whether it performs as well in untrained hands remains the open question the planned home-based study is meant to answer.
Subject of Research: Computer vision analysis of BinaxNOW COVID-19 rapid antigen tests to quantify viral load and predict time to a negative result
Article Title: Seeing beyond positivity: computer vision approach to decoding BinaxNOW COVID-19 tests and forecasting negative results
Article References: Greenleaf, M. N., Parsons, R., Shea, C., Nguyen, P.-V., Waggoner, J. J., Rao, A., Bassit, L., Bowers, H., Sabino, C., Schinazi, R. F., Arconada, S., Mannino, R. G., & Lam, W. (2026). Seeing beyond positivity: computer vision approach to decoding BinaxNOW COVID-19 tests and forecasting negative results. BMC Digital Health, 4(1), Article 63. https://doi.org/10.1186/s44247-026-00287-4
Image Credits: AI Generated
DOI: 10.1186/s44247-026-00287-4
Keywords: COVID-19, rapid antigen tests, lateral flow assay, computer vision, BinaxNOW, viral load, smartphone app, cycle threshold, diagnostics, machine learning, RADx, public health
News Source: Blake Davidson. (October 29, 2026). Smartphone computer vision reads COVID tests and forecasts negative results. Scienmag.
