Fraunhofer MEVIS hat dies direkt geteilt
Our new medical image foundation model is now published in Nature Computational Science: "Overcoming data scarcity in biomedical imaging with a foundational multi-task model" 🎉. We developed a supervised foundational training strategy and pretrained a model (UMedPT) on diverse medical imaging tasks from #radiology and #pathology. Key findings: - UMedPT outperforms ImageNet pretraining & previous state-of-the-art models 💪 - Maintains performance with only 1% of training data for in-domain tasks 📈 - Enables strong results with 50% less data for out-of-domain tasks 🔍 - Extracts highly transferable features validated across multiple clinical centers 🏥🏥 By efficiently leveraging limited data with multi-task learning and gradient accumulation, UMedPT is ideal for #AI in data-scarce medical domains. Thanks to the hard work by Raphael Schäfer, Till Nicke, Fabian Kiessling and after a book worth of peer review comments, UMedPT is now public and ready to try out. Thanks to the colleagues at Fraunhofer MEVIS, RWTH Aachen University, Medizinische Hochschule Hannover, and Universit��t Regensburg and to the constructive peer reviewers! We will continue to further develop the model towards specific applications in cancer research and interactive AI training. We are looking for further application ideas, e.g., in rare diseases or personalized medicine and would be happy to collaborate with clinicians and researchers in these areas! Full paper: https://lnkd.in/eWdyt__n (open access) Model / source code: https://lnkd.in/e95q88_D #OpenSource #ArtificialIntelligence #DeepLearning #DataScience #FoundationModel #RareDiseases #PersonalizedMedicine