About Our Research
Retrospective Re-assessment
In July 2026 we went back and re-examined our own earliest experiments with what we know now. Some of them do not hold up. This page explains what we found, which findings we have withdrawn, and the practices we adopted so the same mistakes do not recur.
This research program started with no lab, no research team, and no formal training in medical AI. It started with a reason to learn quickly and a great deal of reading.
That is a real limitation, and we would rather state it than have it inferred. Our earliest experiments, published in 2025, reflect what we understood at the time. Our current work reflects what we understand now, which is considerably more. Both are on this site. We think the distance between them is worth showing rather than hiding, and the honest way to show it is to say clearly which of the early results no longer stand.
Our first two experiments were measuring something that was not there
Experiments 1 and 2 were anomaly detection studies. The idea was to train a model on normal fetal ultrasound images and see whether it could flag abnormal ones. The dataset we used was organized into folders labeled normal, benign, and malignant, and we took those labels at face value.
They did not mean anything. Checking the images themselves, we found that the normal and benign folders contain the same 335 images, byte for byte, filed twice under opposite labels. Half of the malignant folder is not ultrasound at all: those files are outline drawings, black frames with a single white ellipse, which are the tracing guides that came with the original dataset.
So when our models appeared to separate normal from abnormal, what they were mostly doing was telling an ultrasound photograph apart from a line drawing. That is not a medical finding. The underlying images are fetal head scans from a public dataset that contains no pathology labels of any kind.
We described our training data as much larger than it is
Separately, we had been describing our heart detection model as trained on roughly 1,800 images. The real number is 267. We had been counting one dataset as if it were two, and we had counted a folder of 1,500 files as 1,500 images, when those files were 300 images plus their annotation files. The models themselves were built and measured correctly. Our description of how much data stood behind them was wrong, and that matters, because the amount of evidence behind a result is part of the result.
We could not tell whether our experiments were detecting anything
Our heart detection experiments score models on about 53 images at a time. That is few enough that the normal variation between rounds may be larger than the improvements we were trying to measure. We are now running a dedicated experiment to establish the smallest difference our testing can reliably detect, before spending more effort on improvements we might not be able to see.
Every practice below exists because of a specific way we got something wrong.
- We write experiments down before we run them. The hypothesis, the measurement, and what would count as failure are all recorded in advance, so we cannot decide after the fact that a disappointing result was really a success.
- We publish results that did not work. Our domain pretraining experiment found nothing, and we published it in full.
- We never report a number on its own. Every measurement gets a range, because a single figure with no sense of its uncertainty is close to meaningless on datasets this small.
- We open the data before we trust it. We now check datasets by comparing the actual image content rather than relying on file names or folder labels. Every finding on this page was caught that way, and none of them would have been caught otherwise.
- We correct in public. Corrections stay visible on the pages they affect instead of being quietly edited away.
Experiment 1: Fetus Ultrasound Anomaly Detection
Findings withdrawn
Performance figures withdrawn July 2026. The dataset labels turned out to carry no clinical meaning. Kept online as a record of method.
Experiments 3 to 12: Segmentation series
Unaffected
These used the FOCUS dataset and are not touched by the labeling problem above. Their limitations are the ordinary ones we describe on each page.
Real-Time Fetal Heart Detection (our current work)
Active, with corrections
Corrected in July 2026 to state the true size of our training set. The work stands; our description of it needed fixing.
Why we are telling you this
The simplest option was to delete two pages. Nobody had asked about them and nobody would have noticed. We are keeping them up, with their conclusions withdrawn and the reasons stated, because we are asking clinicians and families to take our current work seriously, and that request is only worth anything if we are equally willing to report the work that failed.
We would rather be an organization that finds its own errors and says so than one that looks flawless. If you find something we have missed, we want to hear about it: info@theodoresheartinitiative.org.