In clinical ultrasound imaging, reconstruction algorithms commonly assume a spatially uniform propagation speed, conventionally set to 1540 m/s, despite the heterogeneity of biological tissues. This approximation can introduce artefacts and geometric distortions in the presence of acoustic heterogeneity, limiting the accuracy of the images obtained. The spatially variable estimation of sound velocity is therefore a key element in improving the quality of ultrasound imaging. The aim of this thesis is to develop a deep learning-based method for estimating two-dimensional sound velocity maps directly from multi-channel radiofrequency ultrasound signals acquired at different insonation angles. To this purpose, a simulated dataset was generated using two-dimensional full-wave simulations in pulse-echo configuration, using the k-Wave toolbox in MATLAB.