#RadioRAG introduces real-time data retrieval to support the accuracy & factuality of LLMs in radiologic diagnosis https://t.co/9o8PLjuwNY @starasteh @DanielTruhn @laim_uka #LLM #LLMs #NLP

RadioRAG retrieves context-specific information from #Radiopaedia in real-time https://t.co/9o8PLjuwNY @starasteh @DanielTruhn @laim_uka #LLMs #radiology #AI

#LLMs using radiology retrieval-augmented generation (RadioRAG) showed variable performance in answering case-based radiology questions https://t.co/9o8PLjuwNY @starasteh @DanielTruhn @laim_uka #Radiopaedia #radiology #AI

Very happy to share our recent article in Nature Cancer on AI Agents for Decision Making in Cancer :) @jnkath

Exciting start to the #AIFOROncology Congress in Milan! First session delivered insightful lectures on data-driven models and federated learning – paving the way for innovative oncology solutions! Stay tuned for a great 2-day experience. #OncologyAI #DigitalHealth

🚨 New Paper Alert! 🚨 We've discovered a major vulnerability in medical large language models (LLMs): they're highly susceptible to targeted misinformation attacks. This could have serious implications for healthcare AI! @DanielTruhn @jnkath Full paper:

Today, we hosted the first conference on LLMs in medicine at @katherlab / @tudresden_de, chaired by @IsabellaWies 🤩 We are not just riding the hype train🚆 - we are working hard to provide scientific & clinical evidence for benefits and limitations of LLMs in healthcare

This review explores the transition of deep learning in radiology from laborious fully supervised methods to more scalable weakly supervised methods. @laim_uka @ekfzdigital @jnkath @katherlab @danieltruhn @LeoMisera @FranzesGustav https://t.co/MuCmWg0D1d

Happy to share our new preprint: "RadioRAG" LLMs 📚 + online RAG (=real-time data🌐) = 📈diagnostic accuracy in radiology questions • Accuracy boosts of 2%-54% • Smaller models performing close to bigger models such as #GPT4 📖: https://t.co/cOum2RhNt4 @laim_uka @DanielTruhn

Resources to start: I recommend starting with https://t.co/EEF7BZgOMW and playing around with the git repository #RadAIChat

Challenges with BERT: Setting up a BERT LLM usually requires technical expertise. Many GPT models (e.g., ChatGPT) are accessible via a browser-based user-interface – less so for BERT models. #RadAIChat

Challenges with BERT: As with any LLM: regulatory approval is a problem. Without it, you expose yourself to legal risks when using LLMs in actual clinical practice. #RadAIchat

Challenges with BERT: It might be necessary to fine-tune the model to the medical domain, see the excellent work by my colleague @K_Bressem: https://t.co/EEF7BZgOMW #RadAIchat

Applications of BERT for Radiology: Protocol assignment, prioritization, and many more… https://t.co/6czS3CSKQg #RadAIchat

Applications of BERT for Radiology: extracting key findings from radiological reports. #RadAIchat

Applications of BERT for Radiology: Identify and correct speech recognition errors: https://t.co/35oyCFmj1Q. #RadAIchat

Applications of BERT for Radiology: Provide outcome information based on radiological reports https://t.co/RuAqON8st0 #RadAIchat

That being said – GPT architectures are also really good at classification:

If you want to generate human-like text – use GPT. If you want to extract information from an existing text – try BERT. #RadAIchat

BERT is good at classification of text, GPT is good at generating new text. #RadAIchat