Artificial Intelligence Helps Radiologists to Focus on Their True Skills
Our world is increasingly getting complex. AI-based computer algorithms can help to better handle challenges in order to focus on the skills that humans are truly good at.
‘There has never been more critical need for improved knowledge management than today,” Eliot Siegel from Baltimore, USA, initially stated. He targeted the increasing difficulty of radiologists’ work: Examinations are getting more and more complex, as the number of sequences and images is exponentially rising. Reporting is nowadays structured in a very complex way to make it eligible for machine reading. Increased utilization of lab data such as genomic and other -omics further complicates analysis and reporting. Even keeping oneself up-to-date regarding scientific developments is a difficult task today. The number of scientific studies is enormously rising: It is estimated that in 2020 it will double every 73 days.
No wonder stress is at an all-time high for radiologists. Burnout is increasing due to radiologist shortage and rising complexity, which downgrades accuracy: Diagnostic errors already outnumber other medical errors by 2-4 times. “I think we radiologists are increasingly feeling like Lucy in the chocolate factory,” Siegel explained by referring to a famous US-TV series from the 1950s in which Lucy struggles to wrap all candies coming down a production line with continuously growing speed. “We need to find ways to handle this speed-up,” he added. “Is there something else than just humans that might be able to come in and help us out?”
The aid of computers could be a great solution, said Siegel. They can help us to reduce typical human errors like using wrong scan parameters and misjudging findings or causes of symptoms. However, computers make errors too, but different ones. So, humans and computers can be very “complementary in working together”.
Different Kinds of Computer Assistance
Today the term artificial intelligence (AI) is ubiquitous. But most people mistake AI for deep learning, which is only one part of AI. Artificial intelligence embraces many different types of computer processing. Deep learning, performed by neuronal networks, belongs to the subdivision of machine learning. Typical neuronal networks learn experimentally, thus somehow similar to human brains, but not in the very same way. “Human brains are far more complex”, Siegel explained. A neuronal network type that is often used in radiology is the convolutional neuronal network (CNN). CNN is capable of image pattern recognition which humans are not very good at.
Computer assisted diagnosis (CAD) is another kind of machine learning that follows a different approach than deep learning algorithms. For detecting, e.g. a certain person within a huge crowd, CAD analyses one face after the other according to certain characteristic features. Developing such an algorithm is a complex task. “It takes months to find it,” Siegel said. Neuronal networks learn much faster by using a large number of cases. It only takes days or weeks to develop an appropriate algorithm. CNNs then do not look at each face like CAD programs, but concentrate on other characteristic image patterns, just like color, shape, etc. and combine them. “You can see neuronal networks like a combination of filters,” the presenter concluded.
Advantages of Deep Learning
Deep learning algorithms are considered to have certain advantages compared to human image analysis. For example, they never get tired, they are typically faster than humans and able to perform multi-tasking. There are already thousands of algorithms being developed that show promising early results. Deep learning is not only capable of image analyses but also of image processing. It helps to reduce imaging time or radiation dose or contrast agent dose while improving image quality at the same time.
However, deep learning also has its weaknesses. There are still many challenges that need to be overcome. While CNNs are good at finding certain patterns in the image, they can’t always reliably identify what is ‘wrong’ in an image, such as pathologies are. Furthermore, the algorithm’s accuracy can easily be disturbed by certain image characteristics resulting in wrong results. Another issue of deep learning algorithms is their testing and approval, as the algorithms processing is still difficult to comprehend. Saliency mapping – a kind of image segmentation that shows each pixel's unique quality – nowadays can help to understand how the algorithm works within their black box. Another issue is that deep learning algorithms require thousands or even millions of imaging data to derive meaningful patterns of information. Obtaining such large data sets is not easy said Siegel: “I don’t have those in my PACS.”
“A Complementing Digital Friend”
Despite these challenges, Siegel is convinced that within the next ten years, machine-learning algorithms will increasingly be implemented in the radiologist’s daily routine. Siegel expects algorithms soon to be available on app-store-like platforms and probably even rule out PACS as the image delivery platform. Nevertheless, these algorithms will hardly replace humans. “Radiologists judge, explain, teach, discover, lead, console, explore, create, and dozens of other things computers can’t even begin to do,” Siegel concluded.
Presentation Title: The impact of AI technologies in patient care: advantages and limitations
Speaker: Eliot Siegel, Baltimore, USA
Date: Wednesday, February 27th, 2019
Session Code: A-0126