Today, physicians can get detailed views inside the human body, thanks to imaging techniques such as ultrasound and computer tomography (CT). Moreover, they can gain precise information about a patient’s condition by means of blood tests and histological findings. And recently, genetic analyses have been added to this armamentarium. All told, these developments are expected to lead to better and, above all, more personalized treatments for patients. However, the flood of data also has its drawbacks. It’s becoming increasingly difficult for doctors to maintain an overview and suggest optimal treatments for their patients on the basis of the latest medical findings. A good example of this is cardiology, where the number of cases is rising as quickly as the amount of diagnostic data and the knowledge about the causes of illnesses and how they can be treated.
AI: Spearheading Improved Patient Diagnostics and Treatments
Medicine is reaching a turning point: Experts believe that in just a few years, artificial intelligence will greatly help doctors make diagnostic decisions and plan treatments. In addition, patients are expected to benefit from increasingly personalized approaches and better care.
Continuous Learning from Patient Data
That’s why systems are needed that can reduce the burden on physicians by deriving precise information about a patient’s condition from large amounts of data and then using that information to suggest promising treatment options. “We don’t need more doctors; what we need are smart solutions,” says cardiologist Ronak Rajani from King’s College in London. Such solutions are becoming better and better because they continually learn from treatment results. This is not only the case in cardiology – other areas such as cancer treatment can also benefit from the use of computers to integrate and analyze patient data so that smarter decisions can be made. The treatment of lung cancer is a good example of this. “We have a flood of data about such cases,” says Bram Stieltjes from University Hospital Basel in Switzerland. However, the interpretation of PET images (positron emission tomography) takes a lot of time and is prone to error. Help could be provided here by an automatic evaluation system that, among other things, analyzes PET images and automatically supplies users with reports.
Machine Learning that Imitates Human Learning
Such systems are made possible by the use of artificial intelligence (AI). The idea is to have computers imitate typically human abilities such as learning (machine learning). Researches have been working to make this vision a reality since the 1950s and they have recently achieved some spectacular successes. AI-based voice assistants are already common and artificial intelligence will soon help autonomous vehicles transport us to our destinations without the need for any human intervention. The use of AI has also produced headlines in medicine. For example, researchers from Heidelberg University in Germany reported in May that a computer program detected malignant melanomas better than 58 dermatologists from 17 countries.
Siemens Healthineers has been working on machine learning since the 1990s and now has more than 400 patents in this field. There are two reasons why AI is just now beginning to transform medicine: Thanks to Moore’s law, computing power continues to increase exponentially so that complex AI algorithms can now be inexpensively integrated into medical equipment. In addition, the digital transformation of the healthcare sector has produced huge amounts of data that AI can learn and draw conclusions from. However, human beings first have to process the data before it can be used to train algorithms. This processing is done by leading experts who analyze findings and annotate them with additional information. To this end, Siemens Healthineers is working together with the best medical experts in the world. In this way, doctors and their patients around the globe are benefiting from the knowledge of top physicians.
Neural Networks Learn from Training Data
The use of machine learning shows particular promise in imaging. Here, the key ingredient is deep learning, in which a computer program imitates the operation of the human brain. In this process, an external layer of artificial nerve cells accepts data such as CT images and forwards it to an inner layer of neurons until the result of the data processing (such as the identification of a tumor) can be called up at the output layer. The network learns how to perform this task through the use of training data. At first, the system is presented with data whose interpretation it already knows. It then gradually adjusts its internal networks in such a way that it achieves the same results. Once this learning process has been completed, a neural network will provide the correct interpretation even when presented with fresh data. Siemens Healthineers has more than 80 patents in this field alone.
When analyzing images, the various layers of a neural network proceed step by step, like those in the human brain. They first recognize fundamental properties such as corners or edges before going on to identify more complex patterns and, ultimately, complete objects. For example, Siemens Healthineers has developed ALPHA (Automatic Landmarking and Parsing of Human Anatomy), which is an algorithm for its syngo.via 3D diagnostic software. The algorithm automatically recognizes anatomical structures and independently numbers a patient’s vertebrae and ribs. Siemens’ ACUSON S2000 Prime, an ultrasound scanner, also contains artificial intelligence. Its integrated software automatically recognizes parts of heart valves (such as the cuspis and the edge) and draws the user’s attention to problems such as the fact that a heart valve may not be closing properly. The algorithm also supplies up to 80 measurement values that provide information about the size and shape of an appropriate replacement valve.
Siemens Healthineers is also using AI methods when planning heart operations. More specifically, one of the unit’s software programs, CT TAVI Planning, is used when a patient needs a new aortic valve. When analyzing CT images of the chest, the software automatically extracts the aorta and ignores other parts of the body. Moreover, it combines individual CT layers to generate a three-dimensional image of the aorta that makes plaque visible. This information enables doctors to determine precisely which kind of replacement valve would be best suited to a patient before they perform minimally invasive surgery for a transcatheter aortic valve implantation (TAVI). The information also tells them where the valve should be placed, how big it should be, and how it should be shaped.
Doctors Will Remain Indispensable Despite AI
“Our research and development activities are permeated by AI technologies,” says Walter Märzendorfer, Head of Diagnostic Imaging at Siemens Healthineers. For Märzendorfer, the collection and analysis of patient data will make it possible to find the best treatment for each individual patient. Other experts also consider AI to have great potential. According to a survey conducted by the magazine The Economist in 2017, 54 percent of the executives in medicine believe that the importance of AI in helping to make treatment decisions will increase considerably over the next five years. However, experts also agree that even though artificial intelligence will be able to assist doctors in the future, it will never be able to replace them.