Skip to main content
. 2023 Feb 11;14:788. doi: 10.1038/s41467-023-36102-1

Fig. 1. Overview of the experimental workflow.

Fig. 1

In V-EUS modeling, the generator takes BUS as input and synthesizes V-EUS, and then the discriminator determines whether the input EUS is real. A color rebalance module and a tumor discrimination module are designed to regularize the model. In V-EUS evaluation, the performance of the model is inspected from three aspects: image quality metrics, determining tumor malignancy, and blind evaluation. We design five experiments in this work. a The model is trained on 2001 high-quality US images from the main cohort. b The hold-out 500 high-quality US images are used as an internal test. c We evaluate the trained model on an external multi-center test cohort with 1730 high-quality US images. d The model is further evaluated on a more challenging dataset containing 349 low-quality US images collected from pocket-sized US devices. e In order to analyze tumor depth dependency, all high-quality US images are divided according to different tumor depth intervals, modeling with 15 mm as division thresholds.