AI Learns to Register Two Medical Images

Inspired by humans, we recently trained an artificial agent to perform the image registration tasks through “Supervised Action Learning”, a technique that combines Deep Learning and Reinforcement Learning under supervision. By just showing examples of aligned images, the agent could figure out how to register CT and DynaCT images together with high performance and robustness. Check out our AAAI 2017 paper, or join us at the conference in San Francisco for our oral presentation!

An Artificial Agent for Robust Image Registration“, R. Liao et al., AAAI 2017

When AI learns human physiology…

Designing an inverse modeling approach to estimate the parameters of a computational model of cardiac function can be challenging, in particular when available clinical data is sparse and noisy. Instead, we propose to train an artificial agent exactly like a new user is trained: first learn how the computational model behaves by random and playful experiments. Then let it devise the optimal strategy for parameter identification through reinforcement learning. Applied on several problems, the agent could learn its new task and perform it much more efficiently and accurately than available inverse modeling strategies!

A Self-Taught Artificial Agent for Multi-Physics Computational Model Personalization“, D. Neumann et al., Medical Image Analysis 2016