Paul F. Jaeger
@pfjaeger
Health AI Research at @GoogleDeepMind. Ex- @DKFZ, @KIT, @ELLISforEurope, and @AIatMeta.
We are releasing TxGemma! 💊💙
Developing new drugs can be notoriously slow, expensive and risky. 🔬 Enter TxGemma: a more general, open and scalable set of AI models to help improve this process for researchers and developers. Find out more → goo.gle/3Y864u7
Developing new drugs can be notoriously slow, expensive and risky. 🔬 Enter TxGemma: a more general, open and scalable set of AI models to help improve this process for researchers and developers. Find out more → goo.gle/3Y864u7
Introducing TxGemma, a family of open models specifically tailored for health settings, building on top of Gemma and Gemini. developers.googleblog.com/en/introducing…
🚨 Published in @Nature today: BOTH of our @GoogleAI @GoogleDeepMind papers on AMIE, our research conversational medical AI 🤖 🩺 @EricTopol - heartfelt thanks for sharing this with the larger community :)
2 important papers @Nature today indicate the potential for #AI to improve diagnostic accuracy and overall clinical performance for patient management and interactions @GoogleDeepMind @vivnat @danmcduff @taotu831 @alan_karthi nature.com/articles/s4158…
Excited to share nnInteractive: The nnU-Net moment for 3D Interactive Segmentation! Supports points, scribbles, boxes, & a new lasso prompt for full 3D masks from 2D inputs. Check it out: arxiv.org/abs/2503.08373 github.com/MIC-DKFZ/napar… mitk.org/wiki/MITK-nnIn…
What’s going on @ICCVConference? Forcing all authors to review is one thing. But not allowing authors unable to review to be removed (even before submission!) is brutal.
🌵Presenting our paper on predictive imaging biomarker discovery at #WACV2025 this Sunday! 📄 arxiv.org/abs/2406.02534… Thanks to @lukas_kln, @peterjensen_, @VollmuthP, @pfjaeger and @maierhein for working on this together @mic_dkfz!
Personal update: After more than 8 years at DKFZ Heidelberg, I recently joined @GoogleDeepMind as a Senior Research Scientist in Toronto. I am very grateful for the amazing colleagues and friends in Heidelberg - and excited for what's next :)
Struggling to pick the right explainable AI method for your vision task? 🚀 Check out our #NeurIPS2024 (D&B) paper on LATEC: a benchmark analyzing 17 methods across 20 metrics and proposing a robust evaluation scheme. Led by @lukas_kln. openreview.net/pdf?id=HRkwnZe…
Out in TPAMI! 🥳 Our systematic review on 121 interactive segmentation methods, including a taxonomy structuring the field and guiding users from their custom problem to an adequate method. Work led by the amazing Zdravko Marinov :) ieeexplore.ieee.org/document/10660…
So literally the day after I made this tweet I saw this great poster @iclr_conf by @kc_kahl, @pfjaeger and others. It’s fantastic work on uncertainty estimation for segmentation methods. Paper: arxiv.org/abs/2401.08501 @mervenoyann you might find this interesting :)
Confidence calibration is one of the biggest problems of our time :) It’s a bad word for an ill posed problem. It makes you think of probability but it’s not. Score might be better, but I’m still on the fence 🤷♀️
In 2021, we published nnU-Net addressing an ongoing bias towards novel architectures in medical image segmentation. Now, 3 years later, we observe that this innovation bias prevails, and - despite recent claims - scaled U-Net variants remain SotA, at least on current benchmarks.
Concerning results: We've evaluated Transformer-, Mamba-, and CNN-based architectures, revealing widespread validation flaws in #MedicalImageSegmentation. Many claims of superiority do not withstand strict testing. Time to push for rigor in the field :) arxiv.org/pdf/2404.09556…
It's really hard for me to state in words how important I think these two sister Perspectives are for the future of bioimage analysis. Kudos to all the authors on their extraordinary efforts. It's been a fun ride handling these papers.
🍻EXCITING DAY: Our @naturemethods sister publications on #MetricsReloaded are out🎉. Recommendations: rdcu.be/dysg5 Pitfalls: rdcu.be/dysgX @helmholtz_image @bias_sig @ProjectMONAI @NCT_HD @DKFZ Important for the community: 1/
🍻EXCITING DAY: Our @naturemethods sister publications on #MetricsReloaded are out🎉. Recommendations: rdcu.be/dysg5 Pitfalls: rdcu.be/dysgX @helmholtz_image @bias_sig @ProjectMONAI @NCT_HD @DKFZ Important for the community: 1/
🌐 Join us for a Special Talk on AI and Infectious Disease Outbreak Prediction with Prof. Joacim Rocklöv, leading Germany's first AI lab in this field. Discover how AI can revolutionize disease surveillance and prediction! 🔗 Details & Registration: heidelberg.ai/2024/01/31/joa…