Artificial intelligence is revolutionising technological development in all fields, including medicine. But before it can reach its full potential, we need to make sure it will be able to manage any given scenario. Some diseases are known for their high variability, and AI need to adapt to all --- even for what it hasn’t been trained for.
How can we improve current AI to properly manage the variability of human illnesses? Read in this week’s Helmholtz AI project showcase how researchers at the German Cancer Research Center (DKFZ) and the German Aerospace Center (DLR) are working on this challenge.
Could you introduce yourself, giving your affiliation, area of work, and of course, the project title?
My name is Katharina Fogelberg, I’m a Data Scientist in Digital Oncology at the German Cancer Research Center (DKFZ). For this project, my group is collaborating with Sireesha Chamarthi, Scientifc Associate in Machine learning at the German Aerospace Center (DLR). Together we are working in the PSDAI project, which stands for Patient-specific diagnostic AI-systems via One-shot Domain Adaptation.
In simple words, what specifically is your project about? And, how and why do you think it is a high risk, high gain endeavour?
Using artificial intelligence in the analysis of medical images has shown a great potential in the latest years. Especially for the field of cancer, AI can work as a preliminary step and help physicians make faster and more accurate diagnosis.
However, these systems only perform correctly when the images shown to them in the clinic resemble the images which were used to train the AI. If the images look different, for example through different lightning conditions or cameras, the performance of the AI usually drops. This can lead to misclassifications and potentially negative consequences for the patient when translated into the clinic.
The goal of this project is to develop techniques which reduce this performance drop and make the AI generalize better towards settings which it did not previously encounter. Therefore, we want to combine two research fields: one-/few-shot learning and domain adaptation.
In order for this to be feasible in a real-world setting such as supporting dermatologists in skin cancer classification, the adaptation process of an AI to a new setting needs to occur fast and with as little data as possible. Realizing this requirement is a challenging task and makes the project high risk. Any successful advancements in this direction would result in a more adaptable AI. This is crucial for patients and physicians alike to increase the acceptance and trust in such AI-based diagnostic systems.
How important has the Helmholtz AI funding and platform been to carry out this project?
Thanks to the Helmholtz AI platform our two Helmholtz centers could come together and found a very interesting use case for which we are able to develop new and innovative machine learning methods. The Helmholtz AI funding enabled the close cooperation between DKFZ and DLR creating exciting synergies and ideas. This collaboration leads to an interdisciplinary exchange between two research groups which rarely work together and thereby allows for the integration of novel perspectives and different experiences. We appreciate the possibilities to exchange with interdisciplinary groups and to learn from each other.