A few thoughts on artificial intelligence in medicine

While AI had a few ups and downs since the 1950's, it typically summarizes methods to tackle challenging problems, e.g., in planning and optimization, in robotics, and in pattern recognition and machine learning. On a conceptual level many methods from AI are surprisingly simple. To solve actual problems, they need to be adapted to the application domain, and this is how a large number of smart variants of the general algorithms have been devised.

Considering that AI methods have been designed to solve hard problems, medicine provides a great application domain. As many biomedical and clinical problems and processes are complex and still not fully understood, it is reasonable to study methods that may not be perfect but solve the problem at hand. However, the inflationary and superficial use of the term "AI" as a placeholder for a vague hope that some black box method will be smarter than us and solve all our hard problems automatically still is rather naive. If you spend some time looking at what's behind AI's success stories you will see a lot of algorithms, smart engineering and careful analysis of the problem and the available data.

We have more than 20 years of experience applying AI methods to actual clinical problems. We have used machine learning to predict kidney transplant failure, beams for radiation therapy, and respiratory motion. We have devised heuristics to solve hard optimization problems in radiation therapy treatment planning. We have worked on smart robots. We have experience detecting and predicting forces and motion of objects from various types of sensor and image data. And we have classified image data to identify tissues ranging from cancer to coronary plaque.

To us, AI is much more than just applying convolutional neural networks. If you are interested in learning more, feel free to contact us schlaefer(at)tuhh.de.