A pipeline to generate large-scale, realistic cardiac ultrasound recordings including clinically relevant artefacts

Abstract

Simulated ultrasound (US) data are widely used to develop and validate (machine learning-based) US data processing algorithms. In this regard, the quantity and quality of the simulated US data are crucial. Here, we have developed an US simulation pipeline to generate realistic cardiac US recordings on a large scale. In this pipeline, we used clinical cardiac US scans to sample the echogenicity of the US scattering sites. In parallel, a non-linear US simulator, k-wave, was employed to generate clinical artifacts due to the presence of ribs and lungs, including reverberation and shadowing. The position of the ventricle, the probe, and the simulated artifact data were then spatially registered in order to modify the originally sampled echogenicities. Motion of the myocardial scattering sites was kinematically governed by a stable mechanical heart model (CircAdapt). The resulting dynamic echogenicity map was fed into a fast convolution-based ultrasound simulator (COLE) to generate cardiac US recordings with clinical appearance including artefacts. The generated US data follow realistic speckle statistics and is a potential augmentation tool for machine learning based US data processing algorithms. [Work funded by European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 860745.]