Achieving excellence across diverse medical applications presents significant hurdles for artificial intelligence (AI), demanding advanced reasoning abilities, access to the latest medical knowledge, and comprehension of intricate multimodal data. Gemini models, Google’s cutting-edge AI, stand out for their robust general capabilities in multimodal and long-context reasoning, presenting promising avenues in the realm of medicine.
In a new paper Capabilities of Gemini Models in Medicine, a research team from Google Research, Google DeepMind, Google Cloud and Verily introduce Med-Gemini, a family of highly proficient multimodal models is tailored for medical tasks, boasting the capacity to seamlessly integrate web search functionality and adapt efficiently to new modalities through customized encoders.
The Gemini models, outlined in the technical reports Gemini 1.0 and 1.5, are transformer decoder models augmented with advancements in architecture, optimization techniques, and training data. These enhancements empower them to excel across diverse modalities including images, audio, video, and text.
Med-Gemini inherits the foundational strengths of Gemini models in language comprehension, multimodal understanding, and long-context reasoning. To bolster language-based tasks, the team enhances the models’ ability to leverage web search via self-training methods and introduces an inference-time strategy guided by uncertainty within an agent framework. This approach equips the model to deliver more precise, reliable, and nuanced outcomes for complex clinical reasoning tasks.
Addressing specialized medical modalities not extensively represented in their pretraining data, the researchers employ multimodal fine-tuning techniques. They demonstrate the models’ adaptability to novel medical modalities by utilizing tailored encoders. Med-Gemini models configured for long-context processing excel in analyzing intricate and lengthy modalities such as de-identified electronic health records (EHRs) and videos.
The team evaluates Med-Gemini across 14 medical benchmarks spanning text, multimodal, and long-context applications, yielding noteworthy results:
- State of the art results on clinical language tasks: Med-Gemini optimized for clinical reasoning achieves a remarkable performance of 91.1% on MedQA (USMLE) through an innovative uncertainty-guided search strategy. Performance enhancements are quantified and validated through meticulous re-annotation of the MedQA dataset by clinical experts. Additionally, the efficacy of the search strategy is demonstrated through leading performance on NEJM CPC and GeneTuring benchmarks.
- Multimodal and long-context capabilities: Med-Gemini attains state-of-the-art performance on 5 out of 7 multimodal medical benchmarks evaluated in the study. The effectiveness of multimodal medical fine-tuning and customization to novel medical modalities, such as electrocardiograms (ECGs), using specialized encoder layers is showcased. Moreover, Med-Gemini showcases robust long-context reasoning abilities, excelling in challenging benchmarks such as “needle-in-the-haystack” tasks within lengthy electronic health records or benchmarks for medical video comprehension. Furthermore, forthcoming work will rigorously explore Gemini’s capabilities in radiology report generation.
Overall, these results provide compelling evidence for the potential of Med-Gemini across various medical domains. However, rigorous evaluation remains essential before its real-world deployment in this safety-critical domain.
The paper Capabilities of Gemini Models in Medicine is on arXiv.
Author: Hecate He | Editor: Chain Zhang
We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.
Pingback: Unveiling Google’s Med-Gemini: Revolutionizing Medical AI with Cutting-Edge Capabilities -