NA-MIC Project Weeks

Back to Projects List

AI-based ultrasound imaging simulator

Key Investigators

• Xihan Ma (WPI)
• Junichi Tokuda (BWH)
• Simon Leonard (JHU)
• Laura Connolly (Queen's)
• Haichong (Kai) Zhang (WPI)

Presenter location: In-person

Project Description

We will discuss strategies to integrate AI-based ultrasound imaging simulators in multiple platforms for medical robotics and IGT, including the Gazebo dynamic simulator and PLUS.

Demo video is available at GitHub

Objective

1. Model Improve the model - make the simulated ultrasound image more realistic
2. Architecture Explore the fast way to package the model into applications.

Approach and Plan

1. Model improvement
   • The current model doesn’t take account of tissue attenuation properties. Explore the use of CT segmentation data (or total segmentator).
   • To use the neuron network to accelerate the computation speed.
2. Architecture
   • Create an independent library for CT-ultrasound conversion. This library takes a 2D resampled CT data that is aligned to the (virtual) ultrasound probe, and generate a corresponding simulated ultrasound image.
   • Integration with existing platforms, including Gazebo, Slicer, PLUS (to be discussed with the community)

Progress and Next Steps

• Progress

  1. Built pipeline to generate sound speed map and density map from CT total segmentation.

     

  2. Built pipeline using k-wave to simulate B-mode ultrasound in 3-dimensional, i.e., the ultrasound beam thickness is taken into account. Tissue speed of sound and density are taken into acount in the simulation.

• Next steps
  1. Debug simulation settings, including transducer properties, beam transmit/receive pattern, etc.
  2. Generate sufficient CT(segmented)-to-simulated ultrasound pairs for neural network training. The network will learn the mapping from CT to ultrasound and achieve real-time processing speed

Illustrations

No response

Background and References

No response