Medical potential of AI

At RISE's Department of Computer Science, a research group works with modern machine learning, Deep Learning, which is used in several types of applications of AI - from environmental analysis to medicine. A study in AI and medicine is currently underway. 

"It is about being able to segment brain tumors using advanced image analysis," says Ebba Ekblom, who works in the research group.  

By taking an image, in this case an X-ray of the brain, we want to get information about where in the image there is a brain tumor and where there is not. This is an important part of diagnosis, which was previously done by hand by doctors. With the help of AI, small signals of abnormalities can be found at an earlier stage than humans can. 

"This can lead to a higher accuracy in finding brain tumors, which saves lives!" says Olof Mogren.

Olof, who is a research leader at the Center for Applied AI and the Department of Computer Science, believes that AI has enormous potential in advanced image analysis for a range of medical applications, such as detecting changes in the eye socket and skin cancer. 

3D models of hospital data

In practice, researchers build models using three-dimensional training data from real hospitals. These are made up of voxels, which show the volume of the entire brain. The machine learning model can see light differences and patterns at different positions in the brain, after it has been trained. 

-The AI model looks for patterns and differences in color and light. Deviations can indicate a tumor," says Olof. 

Federated learning

In addition to working on the medical aspect, the goal of the study is to find techniques where the models work without moving hospital data, as the legal and ethical aspect can be problematic. 
Modern machine learning is data-hungry, so ideally you want to collect all the data in one place. A large model with large training data can generally have higher accuracy than a small model. 

-This is where federated learning comes in! We train a model on the small data set at one clinic, then we do the same thing at another clinic. Then we utilize the federated learning framework that synchronizes these models and learns from multiple data sets," explains Olof. 

Is healthcare ready for AI technology?

Many people contact RISE to learn or get input on how they can use AI. If you want concrete help, the researchers start from the ground up, with questions such as what data is available and what conclusions can be drawn from it.  
"It's a dance we do together, and as researchers we don't always know exactly what we're going to come up with," says Olof.

There are sometimes fears and question marks about whether AI should be allowed to interfere in healthcare-can AI miss something or make the wrong diagnosis? 

-You can imagine that! But do we dare refrain from using AI for health applications, when we can achieve higher accuracy and save lives and suffering?" Olof counters.  

Given that AI can in some cases detect more than humans, it could show specific images that are important for a doctor to look at. Due to legal aspects, it will probably require a doctor to make the final decision. 

"AI doesn't have to replace a doctor, it can be a collaboration where AI and doctors complement each other," Ebba thinks.  

More application areas in medicine

It has taken time for AI for image analysis to be used in medical applications, and here we will only see more and more amazing results. The goal is to find early signals that give a better chance of treating diseases. For the RISE research group in Deep Learning, several medical projects and federated learning await in the future. Sound analysis is on the to-do list, which can help to detect problems in heart and lung sounds or voice abnormalities in diseases such as Parkinson's disease. 

"I think we will come up with many exciting and fun results in the future," Olof concludes.

RISE explain:
 

• Machine learning

A method in AI where computers are trained to learn rules, algorithms, to solve a task.

• Deep learning

Deep machine learning is a subcategory of machine learning, a technique that uses artificial self-learning neural networks. 

• Distributed learning/Federated learning

A learning technique where network nodes utilize local data on multiple devices without exchanging it. It allows several actors, such as hospitals, to build a common machine learning model without sharing data. 

• Federate

Federated means coordinated, and is used in various contexts in the IT industry, such as coordinated databases or coordinated identity management. 

• Voxels 

Three-dimensional pixels, which build up an image in 3D.