Mélanie Bernhardt

Master Thesis
Supervisors: Dr. Valery Vishnevskiy and Prof. Orcun Goksel

Leveraging Simulated and Experimental Data in Image Reconstruction of Sound Speed using Deep Variational Networks

Nowadays, X-ray is the gold standard for breast cancer imaging. However, it does not always succeeds to detect tumors and gives rise to a high number of false alarms. On the other hand, local Speed of Sound (SoS) is a biomarker succeeding to differentiate malignant from begign tumors. Moreover, it has been shown that the local SoS can be obtained from time of flights of ultrasound (US) waves using a standard handheld US imaging probe. Reconstructing the local SoS requires to solve an ill-posed inverse problem. One promising method presented in the literature to solve local SoS reconstruction problems are Variational Networks (VN). Past experiments demonstrate that such networks achieve good results on simulated data. However, they also show that the reconstruction quality on in-vivo data is limited by the generalization ability of the network with respect to the domain shift arising between simulated training data and in-vivo data.