AI in medicine: Machines do not learn like humans

For years, medicine has been exploring AI techniques aimed at easing physician workload. While computers may not have the medical expertise and skills obtained through years of study, they can recognize patterns and specific features in datasets and draw deductions. Though this does not make them as valuable as doctors, it does, however, make them faster and more reliable when it comes to certain tasks.

In this MEDICA-tradefair.com interview, Dr. Markus Wenzel defines the term machine learning, explains how far AI solutions have come in today’s medical practice and ventures a guess as to how they might evolve.

Dr. Wenzel, what do the terms machine learning and deep learning actually mean?

Dr. Markus Wenzel: The term machine learning suggests that a computer sets its own goals and learns something. In my view, that is not quite the case. It actually refers to the process in which a researcher feeds data into the computer aimed at teaching it something. That’s why I believe machine teaching would be a better term for this process.

Irrespective of the term, machine learning is seen as the path to artificial intelligence or AI in short. However, this is only true to a limited extent. In this case, you do not search for methods to replicate and artificially create intelligence. Instead, you look for ways to teach computers to solve a cognitive task as quickly and efficiently as possible and do at least as well as a human being. The method used for this purpose does not necessarily have to be deep learning. It can be any other way of drawing conclusions from data, albeit deep learning is a highly efficient method.

What tasks can AI take over for physicians and perhaps even solve better than they can?

Wenzel: Generally speaking, this refers to duties and responsibilities where physicians systematically have weaknesses because the tasks are repetitive and humans tend to quickly become tired. Or it can be tasks that do not take full advantage of the knowledge and skills of a medical specialist. During a typical day, doctors look at many findings and test results. A radiologist, for instance, reviews many images with a specific task in mind, albeit most images generally only show normal results. These images are recorded based on a suspected diagnosis, though an entire stack of images often only shows results on a handful of images.

This type of time-consuming search task is one of the cognitive abilities where computers can learn and excel in, meaning they review large volumes of data, locate possible areas of concern and presort for the physician so that he or she no longer has to look through all the images.

Are there any products on the market that already utilize deep learning methods?

Wenzel: For the past 10 to 15 years, the U.S has utilized medical devices that use algorithms to analyze certain features on images, which the computer then evaluates and provides a second opinion for the physician. To some extent, the computer also solves the detection task. That being said, these algorithms were not trained via deep learning, but with other types of machine learning, which causes them to no longer be as popular as they used to be.

Deep learning significantly improves algorithm performance. This is made possible thanks to a fundamental change in approach. Previously, we taught computers to look at images the same way we would look at them. This limited our options because – after all – our eye is only a limited "detection machine" so to speak: We look for specific patterns and features that we are able to perceive and detect. If we ask the computer to look for the same features, it adopts the good traits of human beings, while it simultaneously also takes on human shortcomings of course.

Meanwhile, a computer is able to detect other features, which is why it has more opportunities when it can choose the important features by itself. This is why deep learning is often also called feature learning because the computer picks the features it deems important in the images to solve the task we humans have asked it to manage.

What are the objectives of the Fraunhofer MEVIS when it comes to AI research?

Wenzel: We want to use AI methods to ease the workload of physicians. We include physicians in this work and collaborate on finding AI solutions. We also design our tools to offer a certain level of transparency to help physicians understand why computers come to specific decisions and reveal diagnostic findings – or not.

Transparency and understandability of AI solutions are important goals. It promotes acceptance. This is also encouraged by the fact that we let physicians take an active role in these processes, allowing them to personally define the tasks they want computers to handle and determine where and what type of computer assistance they need. Should the computer only flag findings or should it also analyze them? Should the computer suggest treatments or is this something the physician would rather do himself?

What role will machine learning play in medicine in the future?

Wenzel: I suspect there will be many point solutions, each addressing a specific problem, tailored to a specific type of image or designed to recognize a specific disease. Many institutions and research groups are currently pursuing this route.

What’s more – and this is something that will primarily take place in the background – large medical centers and medical device companies will start to process complex data scenarios from a variety of data sources. This would not just pertain to images – they are easily obtainable via the clinical information systems – but would also include all the additional information such as medical reports, radiology reports and laboratory and pathology reports. However, it will still take some time before all information about a specific patient and point in time can be automatically collected from the information systems and processed to where computers can learn from it.

Having said that, it will certainly signal yet another radical change in medicine if computers in the future are able to review a patient’s entire medical history and all images, thus allowing them to track and predict the patient’s journey through a hospital or the entire healthcare system.