Deep-Learning High-Dynamic-Range Ultrasound

In recent years ultrasound imaging has achieved an increasing acceptance across medical specialties. For this reason new techniques keep being tested in the field. Among these techniques we found High Dynamic Range (HDR) imaging where the range of luminosity levels is augmented by combining multiple expositions of a scene. Current ultrasound techniques present limitations that are not compatible with traditional implementations of HDR imaging. In this paper, we asses the use of a deep learning (DL) neural network (U-net architecture) on predicting HDR values from low dynamic range (LDR) input images. In addition, an image acquisition pipeline to create the data set from which the network was trained is described. We demonstrated that this type of networks can be trained to predict HDR out from a minimal number of input expositions, while the obtained results showed to be comparable with more traditional approaches.