Domain adaptation for contrast-agnostic CT volumetric kidney segmentation

The efficacy of segmentation models for CT volumes is limited to the contrast phase they were trained on and often do not work for the non-contrast images. We introduce a domain adaptation approach leveraging a single latent space discriminator to train a robust segmentation model for segmenting CT volume irrespective of the contrast dose. Our model is trained on two publicly available non-contrast and arterial phase image datasets, and validated on both public and private datasets. Evaluation of internal and external tests demonstrates improved segmentation quality while leveraging less data than baseline models.