Medical ultrasound is widely used in clinical practice for its safety, portability, and low cost. Conventional ultrasound imaging mainly relies on B-mode images, which provide qualitative information about tissue structure while assuming a homogeneous acoustic medium. Specifically, the assumption of a constant sound speed of 1540 m/s leads to image distortions, phase aberrations, and defocusing artifacts in heterogeneous tissues. Consequently, quantitative ultrasound methods now aim to estimate spatial maps of sound speed as a robust and clinically relevant biomarker. However, estimating sound speed from raw ultrasound radiofrequency (RF) data is a challenging inverse problem. It is highly nonlinear, sensitive to the acquisition geometry, and intrinsically ill-posed, especially in realistic limited-view settings.