Ryutaro Tanno

Ryutaro Tanno

Ryutaro Tanno

Hi, I’m Ryu. I’ve spent the last 10 years building AI systems for healthcare and biomedical applications. I’m most passionate about creating AI that empowers human experts and transforms their work in ways that are lasting, scalable and joyful.

At Google DeepMind, where I was a Staff Research Scientist until recently, I worked on foundational research and real-world deployment of LLM-based systems for healthcare. Most recently, I led research on AI co-clinician, Google DeepMind’s new initiative to reimagine how care is delivered.

Prior to that, I co-led multimodality for Med-Gemini and was a co-technical lead of AMIE, Google’s research programme on conversational diagnostic AI, helping the team take it from the first text-only diagnostic study to multimodal consultations, specialist care, disease management and a prospective study in primary care. Before that, I led Flamingo-CXR, the first demonstration of human–AI collaboration in radiology report writing.

Previously, I was a Senior Researcher in the Health Intelligence group at Microsoft Research Cambridge with Aditya Nori, where I worked on InnerEye, deployed across the NHS for radiotherapy, and T-Detect, the first T-cell based COVID test. I did my PhD in Computer Vision and Machine Learning at UCL, supervised by Daniel C. Alexander and Antonio Criminisi (Microsoft Research).

Along the way, I’ve been fortunate to work with medical AI start-ups around the world. I was one of the first members of ThinkSono, where I co-developed the prototype of a real-time deep learning system for detecting deep vein thrombosis in ultrasound videos, and a research intern at Butterfly Network, where an algorithm I worked on shipped in their handheld ultrasound device. I’ve also advised SyntheticGestalt and a Terumo spin-out, and currently advise Aillis.

Selected Work

LLMs & Agents for Medicine and Science

Reliable AI: Robustness, Fairness & Uncertainty

Architectures & Representation Learning

  • Adaptive Neural Trees

    Marrying neural networks and decision trees to get the best of both worlds.

  • Stochastic Filter Groups

    Multi-task networks that learn which convolution kernels to share across tasks and which to specialise.

  • Learning Where to Look

    Learning where to sample ultra-high-resolution images before segmenting them, keeping detail where it matters.

  • Learning Better Vision–Language Alignment

    Learning object-level concepts from images and text, and aligning vision–language models with diffusion models.

Full publication list and patents