Our Research
We publish our research openly: methods, results, and limitations. Our program explores how AI can support medical professionals in detecting congenital heart disease earlier and more reliably. Our current focus is real-time guidance that helps capture clear, diagnostic-quality images of the fetal heart, alongside foundational work in cardiac segmentation and anomaly detection. This page is the hub for our experiment write-ups.
All results below are research findings on open datasets under controlled conditions, not clinical performance claims. Our work is a proof of concept and has not undergone prospective clinical validation. See what we do and don't claim.
We register each experiment before running it, recording the question, how we will measure it, and what result would count as a failure. Experiment numbers are registration IDs rather than an order of work, so results appear out of sequence and any gaps stay visible. Browse completed reports or the questions we have committed to and not yet answered.
Published Reports
Registered Hypotheses
Before starting an experiment we write down what we expect to find, how we will measure it, and what result would count as a failure. Those commitments are dated and recorded in our public code repository at the moment they are made, so that we cannot quietly reinterpret a disappointing outcome later as though we had predicted it. Anyone can check that the commitment came first.
Six were registered on 28 July 2026. The five below have not been run yet. The sixth, Experiment 17, has finished and has moved up into Published Reports: it confirmed our own suspicion that our image quality score was measuring the wrong thing, and we retired the figure as we said we would. Their numbers are registration IDs rather than an order of work, so they will not necessarily be published in sequence.
What we expect
We suspect our current way of comparing models cannot reliably detect improvements smaller than about 17 percent, because we score each round on only about 53 images. If so, several of our past comparisons could not have found an effect even if one was there.
What would count as failure
If reanalysing the same data does not meaningfully sharpen our measurements, we will report that no change in statistics rescues a dataset this small, and that our real limit is how much data we have.
What we expect
Our earlier attempt to pretrain on phantom images did not help. We think real fetal ultrasound images may work where the phantom images did not.
What would count as failure
No measurable improvement. Combined with our earlier negative result, that would be strong evidence this approach is not worth further investment, and we would stop pursuing it.
What we expect
We hold roughly 47,500 ultrasound frames that nobody has labelled. We think a model can learn useful general patterns from them and carry that into both heart detection and quality scoring.
What would count as failure
No improvement. Together with Experiments 13 and 15, three failed approaches would tell us our limit is labelled data rather than model design, and we would redirect our effort accordingly.
What we expect
Our heart detector has never been tested on equipment it was not trained on. We expect it to perform worse on an unfamiliar machine than our published figure suggests.
What would count as failure
There is no favourable outcome to protect here. We expect a drop and we will publish whatever it turns out to be, because performance on unfamiliar equipment is the question that matters most for use in under-resourced settings.
What we expect
We think our quality model's GREEN, YELLOW and RED ratings will substantially agree with how trained sonographers rate the same images. No clinician has ever rated an image for this project, so this is currently unknown.
What would count as failure
If agreement is poor, we will retire our accuracy claim and rebuild the model around expert ratings. We will also publish how much the experts agreed with each other, which sets the ceiling any model could reach.