Skip to main content
Serious man in black shirt with tattoo on left arm, standing against white wall, hands in pockets, looking straight ahead.

AI support for medical professionals

Gerardo Hermosillo Valadez | Siemens Inventors of the Year | Lifetime Achievement

Imaging technologies like X-rays, CT scans, and MRIs have been helping doctors for decades by giving them detailed images of bones and tissue structures inside the patient's body. "All of these technologies provide us with image data that’s only helpful with a diagnosis if it can be interpreted correctly. For example, someone has to understand that the fine line in the X-ray image shows that a bone is broken, or notice the unusual mass in the large amount of data from an MRI scan," says Gerardo Hermosillo Valadez from Siemens Healthineers (SHS), Inventor of the Year 2024 in the Lifetime Achievement category. "My department is working on AI processes that can interpret medical images and thereby relieve and support medical staff."

Image interpretation with artificial intelligence

A man in a suit and tie is standing in front of a building with a large window.

Gerardo Hermosillo Valadez is Inventor of the Year 2024: His invention develops anatomical intelligence

People are usually very good at spotting things in pictures. For example, most people will find it easy to recognize the spine in X-ray images – even if the images differ a lot due to different postures of the patient, different angles from which the picture is taken, large or small bones, or healthy vs. deformed spines. Computers can only do this using artificial intelligence – with deep learning – and only if they’ve been trained with many appropriate samples. "This is how artificial intelligence can develop anatomical intelligence so that it can find its way through medical images and correctly categorize the different bones, organs, and tissues" Gerado says. "In the same way, AI can learn to differentiate between normal and abnormal tissue."

Innovations that shape the entire SHS product range

There are many different applications for AI-driven image recognition in everyday life in hospitals and doctors' offices: for example, a system can learn to display the correct image data in response to an instruction like "Show the right shoulder blade." It can also provide automated support for radiological examinations that determines when the injected contrast medium has reached the organ to be examined, and analyze image data for possible pathogenic changes. Many of the innovations that Gerardo has been driving are ubiquitous throughout the entire SHS product range. 

Different AI learning strategies

"Artificial intelligence is only as good as the data it’s been trained with," Gerardo emphasizes. "In order to recognize tumors in an MRI image, it must have first learned from thousands of images of both healthy and sick people that we’ve made available to it." There are three basic approaches to training AI: supervised learning, weakly supervised learning, and self-supervised learning. In supervised learning, the AI is given precise information for each data set: for example, "There’s a tumor in the top right-hand corner of the image." In weakly supervised learning, only a few info labels are given: for example, tumor or no tumor. In self-supervised learning, the data has no labels at all, and the AI learns independently: for example, by hiding parts from existing sample images and "guessing" what they look like.

Person in black shirt standing against white wall, holding a dark object with blurred background.

His invention shapes the entire Siemens Healthineers product range

Serious man in black shirt with tattoo standing against white wall, partially illuminated from right side

Artificial intelligence can learn to find its way through medical images and correctly categorize it.

Almost as smart as an anatomy atlas

Supervised learning is the fastest and most efficient way for AI to learn, but it’s time-consuming and expensive to obtain sufficient training data that takes all relevant aspects into account. Self-supervised learning requires the most learning time, but the AI can work independently without expensive data labelling. "Recently, we’ve been concentrating on the latest self-supervised learning techniques, techniques like those used by ChatGPT," Gerardo says. "That’s given us a significant boost, especially in the field of anatomy intelligence. Our AI is now almost as smart as an anatomy atlas. It can distinguish between about 200 different landmarks (body parts) and locate them in the imaging material.