Domain-agnostic Unsupervised Domain Adaptation Segmentation from 3D Carotid Artery Ultrasound Image

Automatic segmentation of the vessel wall in three-dimensional (3D) carotid artery (CA) ultrasound (US) images can significantly advance the development of cardiovascular disease risk prediction and therapy assessments. Unsupervised domain adaptation segmentation (UDASeg) mitigates the domain shift problem and outputs robust results when the target images are acquired by US machines and imaging parameters different from the training images. Current UDASeg methods focus on target images with domain labels, but domain-agnostic target datasets (DaTD) present more severe domain shift and degrade the effectiveness of UDASeg. To overcome the domain shift induced by the DaTD in 3D CA US image segmentation, we proposed a Domain-agnostic UDASeg (DaUDASeg) method, comprising a Domain-agnostic Target Domain Aligning Module (DaTDAM) and a Cross Pseudo Labeled Supervision (CPLS) in the training stage. Specifically, the DaTDAM decreases domain shift in image appearances and is constructed by a proposed Domain agnostic Style Transfer Model, which adopts a shared network structure and is trained with a Structure Anchor Learning to improve the domain alignment capability in DaTD. The proposed CPLS mitigates overfitting to the source dataset by adding a Pseudo Labeled Target Dataset. This dataset is constructed via similarity filtering and is then used for the final supervised training, providing additional semantic supervision from the target data. The experimental results in a domain-agnostic 3D CA US image dataset showed that our method outperformed the competing methods with 3.0%, 0.09 mm, and 0.57 mm in the DSC, ASSD, and HD95 metrics, respectively. These gains demonstrate improved segmentation performance.