New paper at IJCARS 2019!
Our paper with Santiago Vitale, Emmanuel Iarussi and Ignacio Larrabide on US simulation using CycleGANs was accepted for poster presentation at CARS 2019 in Rennes, France.
Our paper entitled “Improving realism in patient specific Ultrasound simulation using CycleGANs” has been accepted for publication in the International Journal of Computer Assisted Radiology and Surgery (IJCARS)! This is an extension of what Santiago presented at CARS 2019 in Rennes, France.
In this article we present a hybrid approach towards US simulation that combines ray-casting techniques with CycleGANs. The key idea is to use these generative models to alleviate the drawback of current available simulation techniques. In particular, traditional ray-casting based solutions for US simulation usually fail to model complex features such as artifacts or attenuations that are typical from this imaging modality. To overcome this difficulty, we propose to train a CycleGAN using unpaired real US scans and ray-casting based simulations. The model is then applied on simulated images to get more realistic representations. We experimentally evaluated our approach in a user study with a cohort of more than 20 participants, observing that a ResNet based generator is able to significantly improved the appeareance of the images.
We will need more US experts to further validate our approach, so if you are willing to help us please contact Santiago Vitale as email@example.com.
This work was part of a collaboration with Santiago Vitale and Ignacio Larrabide, from Pladema / CONICET, and Emmanuel Iarussi, from UTN and also from CONICET. This is also part of Santiago’s PhD thesis on abdominal ultrasound simulation.