Evaluation of U-Nets for object segmentation in ultrasound images

Ultrasound is a commonly used modality for medical imaging. While this modality has great advantages in terms of safety and cost relative to other imaging modalities, it also has several limitations. Signal-to-noise ratio varies greatly depending on the acoustic properties of the tissue being imaged and the depth of the target structures. In this work, we evaluate the use of deep learning based methods to reconstruct 3D surfaces of general objects imaged with ultrasound. We evaluate three variants of the 3D U-Net with different training scenarios. We were able to train networks to reconstruct three distinct categories of objects relatively well when trained on limited data from each category. However, the performance of the networks did not generalize well when testing on categories of objects not included in the training. We also investigated the effects of employing dual-task autoencoding on generalizability. These results provide a baseline for exploring modifications to the U-Net framework to improve generalizability. A generalizable method could improve visualization for a number of ultrasound imaging tasks.