Improving 3D ultrasound images with AI

What was the problem?

BellyBaby came to us because they wanted to give expectant parents a more lifelike picture of their unborn child from 3D ultrasound images. Low resolution, image noise and elements partly covering the face made it hard to get a clearly readable image. Manual retouching was time-consuming, so the task became to develop a solution that could be automated and built into their service.

What did we build?

We developed an image processing solution based on machine learning models that makes 3D ultrasound images look more lifelike. We made the processing available through an API, so BellyBaby can use it built into their own service.

Comparative model evaluation. We tried several generative AI solutions that produce an image from an image. We examined how closely they follow the shapes and facial features of the original image and what visual quality they deliver. Besides API availability, a decisive factor was the pricing at which the solution could be built into BellyBaby's service: image quality, ease of integration and processing cost all had to be right together.

Balancing image quality and fidelity. Generative processing can also create new details, so we paid special attention to how much the image is changed: besides a more impressive result, following the original facial features was also an evaluation criterion. We examined the level of detail together with the run time, so that the extra computation would bring a real improvement in quality.

A cherished keepsake, without anatomical distortions. A key requirement was that the AI must not create errors such as more than five fingers on a hand. Unfortunately, some of the services we tested produced exactly this kind of result. We did not consider such images acceptable: an extra finger or a distorted hand can turn a cherished keepsake into a frightening experience for parents.

What was the result?

BellyBaby received an AI-based solution that lets them offer their customers a more lifelike image display. Thanks to the API connection, they can provide the image processing as part of their own service.