Wants to reduce debris and interference in MRI

The metrics Anders Garpebring wants to improve concern, for example, blood flow and cell density. These values can then be used for a more accurate plan for radiation therapy.

ImageMattias Pettersson

Anders Garpebring studied engineering physics in Luleå and later moved to Umeå to start doing research. By chance, his research came to be about radiation therapy and magnetic resonance imaging (MRI) – about measuring blood flows in tumors. After his dissertation, Anders went on a two-year post-doc in the Netherlands, during which he worked on a project concerning image enhancement by checking patient-specific variations at high-field magnetic cameras.

Back at Umeå University, his main focus in research is to improve the quality of quantitative measurements made with MRI. The information sought in the images is often indirect in the sense that it needs to be interpreted through models and there are always disturbances in the images in the form of, for example, noise. This means that every image suffer from some uncertainty. Anders wants his research to be able to reduce that uncertainty.

Disturbances creates uncertainty

All images contain noise and other disturbances that risk hiding what you want to see. In our case, says Anders, this contributes to uncertainties in the metrics that we derive from the images.

Radiation therapy can be directed more towards the area where there are the most aggressive tumour cells.

ImageMattias Pettersson

"The uncertainties are exactly what we want to reduce, but we also want to find out how big the remaining uncertainties are. Then the radiation dose can be better adapted for each individual cancerous tumor. The focus of my research is primarily on quantitative images in the field of cancer, but also something towards cognitive ability and aging."

The metrics Anders Garpebring wants to improve concern, for example, blood flow and cell density. These values can then be used for a more accurate plan for radiation therapy. Radiation therapy can, for example, be directed more towards the particular area where there are the most aggressive tumour cells.

The new AI technology can be trained to reduce noise, thus improving the quality of metrics calculated from the images. Because in medicine, high demands are placed on the information in the images.

AI is trained to draw conclusions

"AI, or in our case a special type of AI, neural networks, can be trained based on examples with the hindsight to find patterns themselves. This makes the method very useful in complex situations where it is difficult to give precise rules to, for example, distinguish between benign and malignant tumors in an image material."

"We are in the start-up phase of a project where we will take in images of benign and malignant tumors that AI can train on. Then the AI gets a final exam with information it hasn't looked at before and that's what determines how well it actually works."

One problem with AI technology is that it often requires large amounts of data, which can be both expensive and difficult to access in medical applications. Data in particular will be the most important asset in the future as the use of AI increases. It is easy and cheap to download models from the internet, but, says Anders Garpebring, without good training data they cannot perform the tasks we want them to do.

A bottleneck in breast cancer screening is the large human resources required for review

"There is great potential in this area. Now a computer can take over what has traditionally been done by man. It saves on human resources and, in addition, with AI  it is possible to get more information and analyze more complex data. A good example is in breast cancer screening where an important bottleneck is the large human resources required for review. While 3D images might be preferable for some at-risk groups, it could mean a lot more work when reviewing the images. But with AI methods, that doesn't have to be a problem anymore."

A risk of misjudgment

One risk that Anders Garpebring sees with AI technology is that it becomes too good at what it has learned. For every deviation from what it has learned, there is a risk of misjudgment. Subtle details that aren't obvious at all, can cause an AI model to completely change whatever decision it comes up with.´

"There are ways around that problem, such as the technology telling us that it can't make an assessment because it hasn't seen similar data before."