A rapid diagnostic strip, such as a COVID antigen test or a fluorescent ferritin test, has two lines. The test line (T-line) lights up when the target molecule is present. The control line (C-line) lights up whenever liquid has flowed through the strip correctly. Handheld readers scan the strip into a one-dimensional brightness profile and then decide whether a T-line peak is present and how large it is. The hard cases sit at the detection limit, where a faint T-line peak has to be told apart from a drifting baseline and from peaks distorted by humidity, paper-lot variation, or sample viscosity. Today these cases are handled either by hand-tuned peak finders or by neural networks trained on thousands of labeled profiles. Neither gives a calibrated answer to the clinically relevant question: how likely is it that this faint bump would appear on a truly negative strip?
Flavor physicists face the same problem when they search for tiny matter–antimatter asymmetries in particle decays, and they solve it with two habits. First, they measure the shape of a rare signal using an abundant reference process that passes through the same detector. Second, they report the rare signal through a likelihood-ratio test whose false-alarm rate is checked on samples known to contain no signal. We propose to carry both habits over to strip reading. The C-line is always bright, and the flow and optics that shape it also shape the T-line, so its shape can be measured on each strip and reused as a template. The only free quantity then left for the faint T-line is its amplitude. Fitting that amplitude with a likelihood test, calibrated on blank strips, should lower the practical detection limit at a fixed false-positive rate, need far fewer labels than a neural network, and hold up when the paper or the environment changes. The reason is that each strip provides its own calibration.
Core Claim
The claim is narrow and testable. For fluorescent strip profiles near the detection limit, a template fit in which the C-line fixes the T-line shape and a likelihood ratio decides whether the T-line is present will beat three alternatives at a fixed, empirically verified false-positive rate: threshold rules, wavelet peak finding, and a supervised CNN/U-Net pipeline. The fit wins on detection limit, calibration, and robustness to lot or humidity shift. Our method's contribution is the transfer of shape from the reference line within each strip. Likelihood fitting on its own is a generic tool and not the point.
The measured profile $y(x)$ along strip position $x$ is modeled as
$$ y(x) = b(x;\beta) + A_C\, s(x - x_C;\theta) + A_T\, s(x - x_T;\theta,\delta) + \varepsilon(x), $$
where $b$ is a smooth baseline with a few parameters $\beta$. The function $s$ is a peak shape with parameters $\theta$ (width and tailing from capillary dispersion), fitted mainly by the bright C-line of amplitude $A_C$. $A_T$ is the T-line amplitude, $x_C$ and $x_T$ are line positions known from strip geometry up to a small shift, and $\delta$ is a small, tightly constrained width correction. The correction is needed because the T-line sits upstream of the C-line and sees slightly different dispersion. $\varepsilon$ is noise, with a variance model estimated from blank strips. Presence is decided by
$$ q = -2\ln\frac{\mathcal{L}(A_T = 0)}{\mathcal{L}(\hat A_T)}, $$
and $q$ is compared against a threshold. That threshold is set from the empirical distribution of $q$ on known-negative strips, not from asymptotic theory. Asymptotic theory is unreliable here because $A_T \ge 0$ sits on a boundary.
Relation to Prior Work
The supervised strip reader in Deep learning on lateral flow immunoassay for the analysis of detection data is our main point of comparison. It classifies roughly 4,100 profiles into noise, C-only, T-only, and double-peak classes with a CNN, and it segments about 1,400 double-peak profiles with a U-Net. It reports 99.59% classification accuracy against about 80% for wavelet and direct peak finding. That paper motivates our proposal: it shows that classical peak finding fails on drift and overlap. It also sets the bar we must clear. Our method differs because it produces a calibrated test statistic instead of a class label, and it uses no labels except a pool of blank strips. We also expect the headline accuracy to hide the regime that matters, since the paper aggregates over all concentrations and does not report performance at the edge of the detection limit.
The foundation-model study Interpreting COVID Lateral Flow Tests' Results with Foundation Models constrains our design. It shows that general vision-language models locate the whole cassette well (97.59% IoU for GLaMM) but fail to find the small result window inside it (4.69% IoU), and that they hallucinate, for example calling COVID tests pregnancy tests. We take this as a reason to start from reader profiles, or from a fixed-geometry crop, rather than asking a general model to read the lines.
The physics side supplies the method. The charmed-baryon asymmetry search in Search for CP violation and measurement of branching fractions and decay asymmetry parameters for Λc+→Λh+ and Λc+→Σ0h+ (h=K,π). measures a rare decay mode relative to an abundant reference mode so that detector effects cancel. The C-line plays the role of that reference mode. The BABAR analysis Search for CP violation using T-odd correlations in D0 → K+K−π+π− decays takes the difference of two large asymmetries, about −6.8% and −7.0%, to isolate a tiny one. It shows that differencing works when the nuisance effect is truly shared, and that sharing is the assumption our width correction $\delta$ exists to test. The extended unbinned likelihood fits with kernel-density templates in Measurement of the branching fraction and CP asymmetry for B → D̄0π decays and Measurement of time-dependent CP violation in B0 → KS0π0π0 decays provide the fitting practice: parametric signal shapes, background shapes, and signal yields extracted with honest uncertainties.
On the assay side, An improved procedure based on fluorescence immunochromatography for rapid detection of group B streptococcus from enrichment cultures shows what borderline reading costs clinically. Direct-swab sensitivity was only 34.48%, and 4-hour enrichment reached 90.91% sensitivity with an 11.9% false-positive rate. Some of those false positives are genuine cross-reactive binding, which no reading method can remove, so we treat that cohort only as a motivating regime.
Research Program
We would build the fit as a small, fast library that runs in under a second per profile on reader hardware. Four design questions are studied in turn.
The first is the peak shape family. We compare Gaussian, skewed Gaussian, and exponentially modified Gaussian shapes, and we choose one by goodness of fit on bright double-peak strips.
The second is the width correction $\delta$. We measure how far T-line and C-line shapes actually differ on high-concentration strips, where both lines are bright, and then constrain $\delta$ to that measured spread.
The third is the noise model. We estimate variance as a function of intensity from replicate blank scans.
The fourth is threshold calibration. We set the $q$ threshold for a nominal 5% false-positive rate from one pool of blanks. We then check, on a separate pool of blanks and on blanks from a different paper lot, that the realized rate stays near 5%. This out-of-sample null test follows the validation habit of the physics analyses and is the main safeguard against self-deception.
Baselines
The naive reference point is the standard reader rule. It integrates a fixed window around each line, computes the T/C ratio, and calls a strip positive above a cutoff chosen from the same blank pool so that it hits the same 5% false-positive rate. This is what most deployed readers do. If our fit cannot beat it at matched false-positive rate, the proposal has no value.
The classical signal-processing baseline is continuous-wavelet peak detection with baseline subtraction. Deep learning on lateral flow immunoassay for the analysis of detection data reports about 80% accuracy for it on the four-class task. We run it with its scale parameters tuned on the same training split.
The strongest existing method is the two-stage CNN plus U-Net pipeline from that same paper. On its vendor dataset it reports 99.59% validation classification accuracy and 0.9680 segmentation IoU. We will reimplement it and train it on the same labeled profiles available to us, under the same compute budget and the same hyperparameter search budget as our fit's design choices. Its decision threshold will be recalibrated to the same blank-derived 5% false-positive rate so that the comparison is on detection limit rather than on an operating point. The vendor dataset is not described as public in the material we have. If we cannot obtain it, we will not claim to reproduce the 99.59% figure and will report the reimplementation only on our own data. Beating this baseline means one of two things. Either we achieve a lower concentration at 95% detection probability, at matched and verified false-positive rate, using at most a tenth of its labels. Or we show clearly smaller degradation under the lot and humidity shifts described below.
The key ablation is the same likelihood fit with the T-line shape left free, meaning $\theta$ is fitted separately for each line with no transfer from the C-line. A second, weaker ablation uses a population-average template fixed across all strips rather than one derived from each strip's own C-line. Together they separate three possible sources of any gain: the likelihood machinery, the use of a template at all, and self-calibration from each strip's own reference line.
Why not just…?
Why not just subtract a blank or compute a T/C ratio with a well-chosen cutoff? This is the most serious competitor. When peak shapes are stable, a ratio with a tuned cutoff may capture most of the benefit. Our argument for the fit rests on shape distortion: when humidity or paper lot broadens the peaks, a fixed integration window leaks baseline into the T-line or cuts off part of its signal, whereas a per-strip template follows the change. The shift experiment decides this. If the ratio rule matches the fit under shift as well as in-distribution, we will report that the physics-style fit adds nothing, and that negative result would itself be useful.
Why not just train the CNN on more data? The supervised pipeline already classifies well overall, as reported in Deep learning on lateral flow immunoassay for the analysis of detection data. However, it needs thousands of labels for each assay and outputs a class rather than a false-alarm probability. The label-efficiency curve and the shift experiment show whether more data closes the gap. If the CNN trained on our full labeled set matches the fit on both detection limit and shift robustness, the remaining advantages of the fit are only interpretability and label cost, and we will say so.
Why not just photograph the strip and ask a vision-language model? The benchmark in Interpreting COVID Lateral Flow Tests' Results with Foundation Models reports 64–74% accuracy for GPT-4V, 4.69% IoU on the result window, and heavy hallucination. That is far from the level needed for borderline calls, so we do not pursue this route except for the image extension described below.
Proposed Experiments
The primary experiment uses fluorescent-microsphere strips for a quantitative analyte, ideally ferritin to match the supervised baseline's calibration range of 0–500 ng/mL. We run dilution series concentrated near the detection limit and many blanks, across at least three paper lots and two humidity conditions, and scan them on a handheld reader. The metric is the concentration at which each method detects the analyte with 95% probability, at a false-positive rate verified on held-out blanks. Secondary metrics are calibration curves, the realized false-positive rate on blanks from a lot never used for calibration, and accuracy on the four-class task, for comparison with the published 99.59%.
Three checks would decide the claim. The first is the held-out null test: the realized false-positive rate must stay near 5% on a new lot, and we expect the threshold rule and the CNN to drift more than the fit. The second is the ablation: with shape transfer turned off, the detection-limit gain should mostly disappear. If it does not, the gain comes from generic fitting rather than from using the reference line. The third is the shift test: the fit's detection limit should degrade less than the baselines' when a new paper lot broadens the peaks.
As an exploratory extension, we would apply the same fit to line profiles extracted from LFT-Grounding images using fixed cassette geometry. That dataset reports no profile-level results, so no reported result found on Lacuna for profile-based reading of LFT-Grounding. A null outcome on any of the three decisive checks would be reported as such.
References
- Deep learning on lateral flow immunoassay for the analysis of detection data
- Interpreting COVID Lateral Flow Tests' Results with Foundation Models
- Search for CP violation and measurement of branching fractions and decay asymmetry parameters for Λc+→Λh+ and Λc+→Σ0h+ (h=K,π).
- Search for<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>C</mml:mi><mml:mi>P</mml:mi></mml:math>violation using<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>T</mml:mi></mml:math>-odd correlations in<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>D</mml:mi><mml:mn>0</mml:mn></mml:msup><mml:mo>→</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>K</mml:mi><mml:mo>−</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>+</mml:mo></mml:msup><mml:msup><mml:mi>π</mml:mi><mml:mo>−</mml:mo></mml:msup></mml:math>decays
- Measurement of the branching fraction and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>C</mml:mi><mml:mi>P</mml:mi></mml:math> asymmetry for <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>B</mml:mi><mml:mo stretchy="false">→</mml:mo><mml:msup><mml:mover accent="true"><mml:mi>D</mml:mi><mml:mo stretchy="false">¯</mml:mo></mml:mover><mml:mn>0</mml:mn></mml:msup><mml:mi>π</mml:mi></mml:math> decays
- Measurement of time-dependent <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>C</mml:mi><mml:mi>P</mml:mi></mml:math> violation in <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:msup><mml:mrow><mml:mi>B</mml:mi></mml:mrow><mml:mrow><mml:mn>0</mml:mn></mml:mrow></mml:msup><mml:mo stretchy="false">→</mml:mo><mml:msubsup><mml:mrow><mml:mi>K</mml:mi></mml:mrow><mml:mrow><mml:mi>S</mml:mi></mml:mrow><mml:mrow><mml:mn>0</mml:mn></mml:mrow></mml:msubsup><mml:msup><mml:mrow><mml:mi>π</mml:mi></mml:mrow><mml:mrow><mml:mn>0</mml:mn></mml:mrow></mml:msup><mml:msup><mml:mrow><mml:mi>π</mml:mi></mml:mrow><mml:mrow><mml:mn>0</mml:mn></mml:mrow></mml:msup></mml:mrow></mml:math> decays
- An improved procedure based on fluorescence immunochromatography for rapid detection of group B streptococcus from enrichment cultures