We discuss a general deep-learning framework for the super-resolution of 2D US images (RT-SR) [1] by increasing the image resolution and reconstructing non-acquired scan lines. Applying the framework to US videos with a low spatial resolution and a high frequency (e.g., for the cardiac district), RT-SR generates high-frequency 2D US video with an increased spatial resolution of each frame, thus overcoming the main limits of current US probes, whose spatial resolution decreases as the acquisition frequency increases. To this end, RT-SR trains a neural network to improve the results of the up-sampling to match the target image (i.e., the high-resolution image). The resulting network does not perform the interpolation of the missing lines; this task is already performed by up-sampling. In contrast, the network learns how to transform the up-sampled lines into the target lines. To improve the quality of the up-sampling, RT-SR trains multiple networks; each one specialised to the input anatomical district (e.g., cardiac, abdominal) and its low-resolution image (e.g., 0.5X, 0.25X). This specialisation improves the quality of the up-sampling since we specialise the network to a specific prediction. The execution time of the super-resolution depends on the up-sampling and the network prediction; the prediction is achieved in real-time on standard medical hardware.