黑料福利网

At the start of his PhD trajectory, Tristan Stevens spent time at the Catharina Hospital, working alongside the Radiology department. It was an eye-opener. While undergoing an ultrasound exam may be barely taxing for a patient, the job involves much more for a radiology technician.

鈥淚t鈥檚 really physical work. Pushing, twisting, often under time pressure. Ultrasound exams come one after another,鈥� he says. And with an increasingly diverse patient population, the challenges in the ultrasound room are growing as well. One example, Stevens explains, is performing an ultrasound on a patient with a high BMI.

鈥淪ound waves are reflected differently by fatty tissue, which leads to more noise in the ultrasound images. That makes it harder to quickly find the right location, and the final images are also much less clear.鈥�

First, a quick lesson in ultrasound. Using a special device鈥攖he transducer鈥攈igh-frequency sound waves are sent into the body. These waves bounce off organs and tissues.
The transducer then captures the echoes and converts them into an electrical signal. Ultimately, the system processes the various measured signals into a visible ultrasound image on a screen.

Connecting factor

鈥淢aking an ultrasound image may seem simple. But there are many steps involved before you see the final image. Until now, those steps have been studied and improved separately. In our Signal Processing Systems group, we are taking a completely new approach.鈥�

鈥淲e look at how those separate components can inform and influence each other. Using smart algorithms, we calculate away the disruptive noise, which should lead to more reliable ultrasound images.鈥�

An ideal project for Stevens, a natural connector who enjoys bringing people and information together and dislikes silos. At the start of his PhD, he also helped bring together the 黑料福利网 chess community as a of the student chess association Noesis.

鈥淚 get a lot of energy from working together. In a research setting, that can offer a refreshing perspective鈥攅xactly what you need for out-of-the-box innovation.鈥�

Self-thinking ultrasound device

In recent years, major steps have been made in this field by the research group led by Ruud van Sloun, one of Stevens鈥� supervisors. According to Stevens, they are now close to a first concept of cognitive ultrasound. In this approach, an ultrasound device determines for itself鈥攂ased on smart algorithms鈥攚hich measurements should be taken. Sound waves are then directed toward the location with the most relevant information.

A breakthrough in medical image processing, Stevens emphasizes enthusiastically. 鈥淭o achieve this, we work with so-called deep generative models. This is a form of artificial intelligence that uses deep neural networks to learn the structure of data and, based on that, can generate new content.鈥�

鈥淢uch like ChatGPT predicts which word comes next based on patterns in large datasets, these models can predict which echo signals are most likely, allowing them to suppress noise or reconstruct missing information.鈥�

Dynamics

To train his models, Stevens used specific ultrasound data from collaboration partner Philips Research. The focus was on cardiac ultrasound鈥攂oth because the heart is highly dynamic, which makes it challenging, and because there are many people with cardiovascular diseases. But, Stevens stresses, other forms of ultrasound imaging can benefit from this technology as well.

Stevens flips through his dissertation and shows several ultrasound images of a heart. 鈥淲ith this new model, we see a clear improvement in image quality. This also applies to more complex ultrasound imaging, such as in the growing number of patients with obesity.鈥�

Investing in coding

His research is highly technical, Stevens says. At the same time, he is keen to help pave the way for actual clinical application. To accelerate innovation, Stevens therefore built a new open-source software platform, .

鈥淎round me, I found a lot of expertise, but all in separate pieces. Now I鈥檝e brought together all the ingredients you need for smart ultrasound: models we鈥檝e trained, tools to train your own models on ultrasound data, and methods to determine what you should measure next to create the following image.鈥�

Developing such a platform is a substantial investment, the Electrical Engineering PhD candidate adds. 鈥淎t the university, you鈥檙e assessed on the number of papers you publish. Code or a platform you build is seen as a byproduct. While that is also very important鈥攊t saves everyone a lot of valuable time if, after reading a paper, you can immediately get to work. Fortunately, there is growing appreciation for that.鈥�

Smartphone ultrasound

For zea, Stevens did not use the programming language commonly used within the ultrasound community, but instead relied on state-of-the-art machine learning code in Python. According to him, the major advantage is that new technological developments from computer science can be directly applied to collecting and processing ultrasound signals.

This makes the effect on image quality immediately visible, Stevens explains. And it brings portable ultrasound equipment鈥攁llowing you to make reliable ultrasound images with your smartphone鈥攁 step closer to the quality of a full-fledged hospital scanner. He lists a few more examples. 鈥淪harper ultrasound images in ambulances, even for patients with severe obesity. Or better prenatal ultrasounds in countries with limited healthcare.鈥�

鈥淎 nice bit of legacy,鈥� he smiles. 鈥淲e鈥檝e shown that deep generative modeling is a powerful tool that appears promising for successful application in ultrasound imaging, even for raw data. Many colleagues are already using the new platform. A small building block toward making the ultrasound of the future smarter.鈥�

This article was originally published by Cursor, the independent news site of 黑料福利网. It is not a press release and is protected by copyright. No part of this publication may be reproduced without explicit permission from the Cursor . This article was translated from Dutch to English using AI-assisted tools and reviewed by an editor.