Beyond the visible: The true state of AI in medical imaging
“I have always been passionate about AI,” says Lena Maier-Hein of the German Cancer Research Center (DKFZ). “In my studies at Karlsruhe Institute of Technology and at Imperial College London, I specialized on the topic although a lot of people back then said that it would never really generate real-world impact at scale. But AI, and especially neural networks, simply fascinated me. For my Master’s thesis, I worked on automatic speech recognition using electromyographic (EMG) signals, but I soon realized that I wanted to apply my skills in the field of life sciences. This eventually brought me to medical image analysis.”
Maier-Hein is division head of (Intelligent Medical Systems (IMSY) at the DKFZ as well as managing director of the National Center for Tumor Diseases (NCT) Heidelberg. At NCT, she works with an interdisciplinary team of six directors from various backgrounds, including medicine, biology, and ethics.
“As a trained computer scientist, I mainly bring in expertise in data science and AI,” says Maier-Hein. The tasks are very diverse, ranging from budget responsibilities to the organization of events. For me, the most exciting aspect of the job is the strategic planning. For example, the data science activities of the NCT Heidelberg have so far been conducted in isolation for the different clinical application domains radiology, radiotherapy, surgery, etc. We have recently set up a cross-topic program “Digital Oncology” with its own budget to enable multidisciplinary data science for personalized cancer care.”
Exciting projects are ongoing
In 2020, Maier-Hein was awarded a European Research Council (ERC) Consolidator grant for NEURAL SPICING. “It is supposed to be an individual grant, but it really was a DKFZ- IMSY lab team effort from concept to grant writing to interview preparation.” This project is a favorite of Maier-Hein as it combines multiple fascinating aspects: novel imaging concepts based on biophotonics, AI as the core enabling technique, exciting methodological challenges, and clinically relevant questions. “In the spirit of model-based AI, the primary technical mission is to make spectral imaging quantitative with a learning-to-simulate approach,” says Maier-Hein.
Another favorite undertaking is the Helmholtz Imaging project on AI benchmarking at DKFZ, which focuses on standardizing evaluation and benchmarking practices in the field. Maier-Hein explains the mission of Helmholtz Imaging as aiming to unlock the potential of imaging in the Helmholtz Association — Germany’s largest scientific organization. “DKFZ is one of the coordinating institutions, and my personal role is the coordination of activities in the field of image analysis benchmarking.”
Maier-Hein notes that emphasis is often placed on developing new solutions, but sometimes researchers forget that they are not able to reliably measure scientific progress or to decide on the clinical applicability of a method if their validation is flawed. “Together with academic and industrial thought leaders from all around the world, we are revisiting common practice and will hopefully generate practice-changing impact in the end,” says Maier-Hein.
Maier-Hein's main message for the audience at SPIE Medical Imaging is: question everything! “If we really want to generate patient impact, we need solutions that work in the wild,” says Maier-Hein. “To this end, we need to design every step of an image analysis pipeline very carefully, from the choice of input data to the annotations and the validation strategy.”
In her keynote presentation, Maier-Hein will show that historically grown practices may be suboptimal, thus hindering clinical translation. “Challenging common practices and addressing them — literally ‘re-searching’ — is an important part of science, I think,” says Maier-Hein.
|Enjoy this article?
Get similar news in your inbox