Emma Chen

One of the coolest aspects of this work is how we leverage the MedInterp dataset. Its diversity across 11 tasks and 3 modalities allowed us to train MedVersa in a unique way. By using domain-aware minibatch gradient descent, we constructed minibatches from the same task and imaging modality, enabling MedVersa to optimize for specific tasks and modalities. This approach, combined with the dataset's richness, was key to developing MedVersa's versatility and robustness. 💪📊

We announced MedVersa, a generalist AI that excels in multifaceted medical image interpretation! 🚀🩺 👉arxiv.org/abs/2405.07988 MedVersa has two promising features🧐: 1. Learning from vision and language supervision. This maximizes the flexibility of the framework. Imagine combining the SAM and CLIP into a unified model and the mutual benefits it will bring!🤪 2. Leveraging existing modules/tools. MedVersa functions the LLM as a TRAINABLE orchestrator. This design features a notable level of extensibility, allowing integration with advanced modules/tools.🤩 A great journey with @subathraadithan and @jn_acosta , led by @EricTopol and @pranavrajpurkar Stay tuned for more updates!👀 #artificiallyinteligence

Are you passionate about advancing global healthcare and save lives through data? We're hiring for a full-time role at Harvard where you will play a pivotal role in expanding our international data-sharing coalition. tinyurl.com/y6d5f8zy rajpurkarlab.hms.harvard.edu/maida

Every week, for the past 3+ years, we summarize the important #AI papers in medicine, life science, and healthcare. It's free to subscribe and we've just moved to over @substack doctorpenguin.substack.com w/ @emma_ychen and @pranavrajpurkar

Medical data used to train AI algorithms lacks diversity. Most data comes from a few hospitals in the US. We are launching an initiative to bring you diverse medical image datasets from across the globe: MAIDA. Here's what MAIDA is about. 🧵 rajpurkarlab.hms.harvard.edu/maida

Key deployment consideration for AI + medicine Can we predict when AI models for medical image diagnosis will make incorrect predictions? New paper (#CheXbreak): arxiv.org/abs/2103.09957 w/ @emma_ychen, Andy Kim, @RayanKrishnan, @JinLong19844963 @AndrewYNg @StanfordAILab 1/n