AI — a resource in medical assessment

At RISE, the Center for Applied AI, a computer is working hard on machine learning, a method where the computer learns by following the same algorithm repeatedly.

"Artificial intelligence is about getting a machine to understand things like a human by teaching itself," says Fehmi Ben Abdesslem, senior researcher in intelligent systems at RISE, the Center for Applied AI.

When the computer is satisfied with the learning process, after about two weeks of work and thousands of repetitions, it picks out a three-dimensional model of an adrenal gland, assembled from hundreds of images in just a few minutes. On the computer screen, the outline of the organ is highlighted[1] in red and green, demonstrating in concrete terms the difference between human and computer ability to draw the outline of the same organ.

"Training AI algorithms aims to remove the difference between the red and green markings, so the model thinks the computer is doing as good a job as a human," Fehmi explains.

The pilot project is a collaboration between RISE, Karolinska Institutet and Karolinska University Hospital, with support from Region Stockholm. The idea of the research project is to use AI to more clearly see and detect metastases, which would help healthcare professionals and cancer patients. The adrenal gland, an organ that produces the vital hormone adrenaline, was chosen as a suitable organ to study.

"The adrenal gland is important to study precisely because it is a common location for metastases - of lung cancer, melanoma, and other cancers. In addition, it is an anatomically complex organ that is difficult to measure by hand because of its irregular shape and where small metastases are easy to miss," says Vitali Grozman, senior physician at the Department of Radiology at Karolinska University Hospital.

Fehmi creates a "recipe", called an algorithm in computer language, where the computer is programmed to know what to look at and what not to look at based on set parameters. In this case, finding the limits of the organ! The algorithm must constantly evolve in order for the computer to learn new things on its own, as well as needing materials to practice on and solutions. In this case, it uses images from computed tomography (CT), an examination method using an X-ray machine that has the capacity to generate several thousand cross-sectional images of a whole body. Based on these images, where doctors have drawn the outline of adrenal glands by hand, the AI can be trained. Over time, the computer learns to make its own conclusions, and this is when unexpected results beyond human capabilities can be obtained.

"What we see in the study is that AI can identify the voxels[2] of an organ better than the human hand. At the beginning of our study, the difference between red and green on the image almost always meant that the AI did a worse job than the human, now it is sometimes the opposite. Moreover, we have seen that the computer learns more than just drawing the outline of the adrenal gland, and we can now cautiously say that we can teach the algorithm to find tumors in other organs as well," reveals Professor Magnus Boman of Karolinska Institutet.

Since an increase in the volume of the adrenal gland can be a sign of cancer or metastasis, artificial intelligence would be of great help to radiologists. Both in terms of finding adrenal tumors more accurately, calculating the volume of the organ and streamlining the doctors' work process. This in turn can contribute to faster and personalized treatment, known as precision medicine. The collaboration between RISE researchers, doctors and academia is important, as the results of the training need to be discussed regularly to develop the algorithms. In the long term, the team wants to expand the collaboration and apply their experience to other application areas of AI in medicine.

RISE explains more about machine learning:

Machine learning is about teaching a computer to teach itself. This takes place in a unit known as a GPU (Graphics Processing Unit). It is a component that performs specific operations extremely fast, like an extra super brain for the computer. It is also needed for learning:

1. Examples and solutions (supervised learning)
2. examples without solutions (to guess from the first example)
3. A programmed algorithm (a "recipe")

This is how the computer is instructed to teach itself:

• Take a closer look at this area!
• Do it with better precision!
• Do the same thing but with different organs!
• Put the knowledge into practice!

[1] Red marker is drawn by radiologist, green is drawn by AI.

[2] Voxels are three-dimensional pixels.

Text: Helena Erngård