News

New HANAMI publication presents innovative method for optimising surgical outcomes using reinforcement learning and computational fluid dynamics.

 

Researchers from the HANAMI project have published a significant advancement in using artificial intelligence (AI) and physics-based simulations for nasal surgery planning. The study, titled “Parallel Reinforcement Learning and Gaussian Process Regression for Improved Physics-Based Nasal Surgery Planning“, in collaboration with Jülich Supercomputing Centre, RWTH Aachen University, Kobe University and RIKEN Center for Computational Science, proposes a novel approach to help surgeons plan interventions that are more personalised, precise, and effective.

 

Bringing Science into the Operating Room

 

Nasal surgeries aim to improve airflow and breathing quality. Yet, current planning techniques — such as CT imaging and rhinomanometry — offer limited insight into how air moves through the complex structure of the nasal cavity. The HANAMI team addresses this gap by combining Computational Fluid Dynamics (CFD), which accurately simulates airflow, with Reinforcement Learning (RL) — a branch of AI that learns how to optimise the shape of the nasal airway to improve breathing outcomes.

 

This combination enables not only the evaluation of a patient’s current condition but also predictions of how surgical changes might affect airflow, effectively allowing surgeons to “test” different scenarios before making real incisions.

 

A Smarter, More Efficient Planning Process

 

The study introduces two key innovations that significantly enhance the efficiency of the surgical planning pipeline. First, parallel reinforcement learning is applied to train multiple AI agents simultaneously, helping the system avoid getting trapped in local optima and resulting in more reliable and optimised surgical recommendations. Second, Gaussian Process Regression (GPR) is applied to reduce the high computational cost of repeated CFD simulations. By predicting airflow performance without requiring a full simulation each time, GPR can reduce simulation time by up to 6% without compromising accuracy. Together, these improvements make the planning process faster, more efficient, and bring it a step closer to real-world clinical application.

 

Towards User-Friendly Surgical Tools

 

One challenge the researchers are actively addressing is usability. Currently, generating a 3D model of a patient’s nasal cavity requires manually editing CT images slice by slice, which takes hours. The team is developing a new drag-and-drop interface that allows surgeons to intuitively adjust the 3D model within minutes, saving time and making the tool more accessible for clinical use.

 

Although clinical trials are not part of the current research scope, these improvements pave the way for future real-world applications.

 

Advancing HANAMI’s Mission: Human-Centred AI for Healthcare

 

This work, led by the Jülich Supercomputing Centre (Forschungszentrum Jülich) in collaboration with the RIKEN Centre for Computational Science (Japan), reflects HANAMI’s broader mission: to integrate AI, High-Performance Computing (HPC), and physics-based modelling into human-centred medical innovation.

 

By hiding computational complexity and focusing on practical, surgeon-friendly tools, the team is laying the groundwork for a new generation of open-source technologies to support ENT (Ear, Nose, and Throat) medicine, including applications in surgery planning, disease assessment, and viral risk evaluation.