Pan, Y.-C.; Vienneau, E.; Lefevre, R.; Eagle, S.; Byram, B. “Overcoming Labeled Data Barriers in Deep Ultrasound Imaging.” European Signal Processing Conference, 2024, pp. 770-774.
Deep networks have significantly advanced medical imaging. Initially, they were used for diagnosing conditions by interpreting images (classification). More recently, they are being applied to create the images themselves (estimation). This study focuses on ultrasound imaging and explores a method to address challenges with unlabeled data.
Fig. 1.
In (A), we show our method for resolving domain shift, providing us with the maps GST and GTSbetween simulated and in vivo data and the reverse. In (B), we then train a deep beamformer simultaneously regressing on sims and in vivo proxies. The two data types are both allowed to contribute to the beamformer, but augmented feature mapping is used, so that any aspects of the beamforming that are distinct to the simulated or in vivo data are preserved.