Deep Multitask Learning for Ultrasound Beamforming

In the field of ultrasound image formation, the beamforming stage plays a crucial role in determining the quality of the final image. Its primary function is to enhance signal quality and image clarity. Our research focuses on the challenges involved in developing ultrasound image formation techniques, with a particular emphasis on multitask learning to create switchable and controllable multi-output beamforming algorithms . We have developed a novel weight normalization scheme for multitask
learning, which has proven effective in adapting a trained neural network to handle multiple tasks, such as speckle noise reduction and reconstruction from subsampled signals. This scheme provides enhanced control over task intensity through a method for adjusting the impact of weight transformation. Furthermore, we have designed a lightweight multitask-based deep learning method that not only enhances image contrast but also maintains minimal computational overhead.

M.Sc. student under the supervision of Prof. Israel Cohen.