RWTH Aachen Pioneers Foundational Model for Advancing Biomedical Imaging
RWTH researchers have developed the first foundational model for biomedical imaging.
A research team led by Professor Fabian Kiessling from RWTH’s Institute of Experimental Molecular Imaging and Fraunhofer MEVIS and Dr. Johannes Lotz from Fraunhofer MEVIS has achieved a significant breakthrough in biomedical research.
Together with researchers from RWTH Aachen University, the University of Regensburg, and Hannover Medical School, they developed UMedPT, a pivotal foundational model biomedical imaging. Their work, titled Overcoming Data Scarcity in Biomedical Imaging With a Foundational Multi-Task Model, has been published in the prestigious journal Nature Computational Science.
Pretraining with foundational models, similar to the approach used in ChatGPT, can reduce the amount of training data needed by leveraging existing knowledge. However, foundational models must be trained on large, comprehensive data sets, which are often not available in the biomedical field. To address this challenge, the researchers developed UMedPT and rigorously validated it using X-ray images as well as microscopic and tomographic images. UMedPT employs a specially designed, novel multitasking learning algorithm that can be trained on a diversely labeled set of small to medium-sized biomedical imaging datasets.
“Our results demonstrate that a single network can surpass the state of the art,” says Professor Fabian Kiessling. “In the international SemiCOL competition, UMedPT outperformed all other methods in colorectal cancer classification.” UMedPT also outperformed ImageNet-pretrained models and the best published reference results with only 1 percent of the original training data for tasks with related data in the pretraining database.
Even for tasks with no related data in the pretraining database, UMedPT exceeded the performance of ImageNet-pretrained models and the best published reference results in the field, using 50 percent or less of the original training data.
These advances show that UMedPT offers significant benefits for applications where sample data is limited – a common issue in biomedical research, particularly for rare diseases, according to Kiessling.