My name is LiXinchen, and I am a computer science student. My current research project focuses on Reinforcement Learning robotic ultrasound scanning, and I want to do experiments in a virtual environment before moving to real-world setups.
However, I am not sure how to construct a 3D volume that contains realistic ultrasound information, which is essential for my simulation and robotic control experiments. After some investigation, I found that one possible workflow is as follows:
Use CT images with segmentation labels to simulate a sequence of ultrasound images.
Apply Sweep Compounding to the simulated ultrasound images in order to reconstruct a 3D ultrasound volume.
Use this volume to obtain ultrasound information from arbitrary directions (i.e., any plane) within the virtual environment.
Could you please confirm whether this approach is correct? Specifically:
Will the 3D volume generated through Sweep Compounding of the simulated ultrasound images truly allow for reslicing in any direction, just like a real 3D ultrasound dataset?
Are there any best practices or recommended settings in ImFusion Suite for obtaining the best results when following this pipeline?
If there are example datasets, tutorials, or documentation that could help me implement this workflow, I would greatly appreciate any pointers.
Hi Xinchen,
Ultrasound imaging is, by nature, direction dependent. Think for example of a bone casting a strong shadow behind it, depending on the direction of the sound beams hitting the bone (so probe position and direction) this shadow will have different appearances and directions. Therefore, reslicing a 3D ultrasound volume will never produce perfect results.
If you don’t need to simulate artifacts like shadows, reslicing might be sufficient otherwise you would have to use the ultrasound simulation as part of your virtual reinforcement learning environment.
If you decide to use reslicing these would be my suggestions:
If you have access to a robotic ultrasound system, I would suggest to avoid using the simulation. Use your system to record sweeps and compound them all together to get a 3D ultrasound volume.
Otherwise, we can have a look online if there exists some already available datasets for your target anatomy.
If it is not possible to get real ultrasound data, then you have to use the simulation. The settings of the simulation algorithm are highly dependent on the anatomy and segmentation labels, therefore I would need more information to help you there. The used documentation contains some useful information about Ultrasound Simulation
Thank you very much for your detailed explanation and suggestions — they were extremely helpful.
For now, I will go ahead and try the reslicing-based approach first. If I run into limitations or need higher realism later on, I may explore the simulation option further.
Thanks a lot for sharing the example workspace — this looks very helpful!
I’ll take a closer look at the Hybrid Ultrasound Simulation demo and see how I can apply it to my setup.
Really appreciate your support!
Best regards,
Xinchen
