Generalization of a deep learning network for beamforming and segmentation of ultrasound images

Recently, deep neural networks (DNNs) for beamforming and segmenting plane-wave ultrasound images have been proposed. The promising results obtained so far focus on segmenting anechoic, almost circular structures using one architecture trained on a large dataset. We present a study of DNNs generalizability for beamforming and segmenting structures of various shapes and echogenicity. Three different encoder architectures (i.e. VGG13/16/19) and target images with standard dynamic range (dR = 60 dB, E60) or an automatically determined dR (Eauto) were compared. Field II was used to simulate 6560 images (with hyperechoic, hypoechoic, anechoic and mixed targets) using random bunches of ellipses to generate different shapes for DNN training. The test set included 816 simulated images, 21 images of a phantom (CIRS040GSE) and 24 images of the carotid artery. The DNN architecture has 1 encoder and 2 decoders, for segmentation and beamforming, based on the UNet. Using the VGG19 trained with Eauto images, a considerable improvement was achieved when compared to other architectures, especially when performing tests on experimental data. Overall, the promising results obtained encourage us to further investigate the use of DNNs for beamforming and segmentation, with the aim to improve the performance and generalize their use for specific ultrasound imaging applications.