Experiment Two:

Our second major research effort, "Iterative Improvements," is a methodical investigation into how we can continuously enhance the performance of our AI models for detecting fetal heart anomalies. Following our initial experiment which established a baseline, this phase is dedicated to systematically optimizing our AI’s ability to "see" and interpret prenatal ultrasound images.

The core of this experiment involves a series of controlled stages. In each stage, we introduce only one specific change to our AI model – for instance, modifying its internal architecture or adjusting its learning process. This precise approach allows us to isolate the impact of each individual adjustment on the model’s performance.

Our findings varied significantly with each iteration. For example, the introduction of a standard U-Net AI structure yielded a neutral impact, performing similarly to our initial model. However, when we implemented an Attention U-Net, which helps the AI focus on critical image areas, we observed a positive and substantial improvement in its detection capabilities. Conversely, other modifications, such as more complex image preparation techniques or a specific learning strategy, unfortunately did not improve performance and, in some cases, even led to a slight reduction in accuracy.

Through these rigorous, iterative tests, we are gaining critical insights into which adjustments most effectively advance our AI. This experimental process is vital for building a highly reliable and precise tool that can provide invaluable insights for medical professionals.

Looking Forward

With these initial experiments under our belt, we are very excited to expand our training methodology to encompass larger data sets, containing more specifically pertinent content in order to produce more robust models with a higher level of accuracy on a broader selection of validation images. We will be looking to enrich our data with pre-training meta data and labeling to provide the best possible results.

While we have been able to complete our first experiments locally on existing hardware we will be expanding future experiments out using cloud based resources to leverage more powerful compute to yield better results from larger data sets.

These future experiments will hopefully lead us farther down the path toward a model that can be used to reliably identify when an image of a heart in a routine ultrasound appears to be abnormal in some way and warrants further testing by human experts. The generous donations we have received and continue to come in are vital to our ability to conduct these experiments on costly cloud resources and we are so grateful for the opportunity to explore what is possible with cutting edge AI solutions for early detection of congenital heart defects.

Our heartfelt thanks go out to those who’ve contributed during our seeding stage. Your kindness and generosity have been vital to our success. We truly appreciate your support.

Issue: 1 Posted: 7/23/2025