Experiment Update:

AI Model Development Progress

This month, we've achieved significant progress in our research on the best way to train a computer vision model to detect congenital heart defects in Ultrasound images. In an effort to thoroughly explore the strengths and weaknesses of different training methods on the publicly available datasets we currently have available to us, we ran through a series of 7 new experiments each building off one another to squeeze as much insight as possible out of the results of each experiment.

Our journey began by testing different approaches—from models that learn on their own to those we directly teach using labeled examples. We found that a supervised approach, where we train the AI to identify and outline cardiac regions in ultrasound images, showed the most promise.

The real breakthrough came through systematic refinement. We fine-tuned how the model learns, introduced advanced techniques to help it handle varied real-world images, and most critically, reduced the rate of missed detections by 20%. Because our mission is to improve the ability of a human technician to identify abnormalities for further review, so this improvement is particularly meaningful to our mission.

To ensure our results were reliable and not just lucky, we validated the model using a rigorous testing method called cross-validation. The final model achieved a performance score of 0.70, surpassing our clinical threshold of 0.70 and demonstrating consistent, trustworthy results.

Want to explore more? Dive into the experiment details and discover fascinating insights.