Domain divergence minimization for unsupervised domain adaptation cross-modality medical image segmentation

Unsupervised domain adaptation is becoming increasingly useful in cross-modality medical image segmentation, especially when labels are unavailable. Existing works aim to mitigate the effect of domain shift through image transformation and feature alignment. However, generative adversarial network (GAN) based image synthesis and implicit feature matching methods suffer from complex network structures and unstable training. In this work, we propose a novel capsule variational autoencoder (VAE) framework with a new divergence, aiming to effectively capture both domain-specific and domain-invariant features. Specifically, two capsule networks are employed to extract discriminative domain-specific features, leveraging their ability to capture spatial orientation and positional information. A shared encoder then maps these features into a common latent space, where a prior matching module minimizes the log-ratio gap derived from Hölder's divergence, thereby promoting domain-invariant representations. The latent features are further utilized to reconstruct input images and generate accurate segmentation results. Extensive experiments are conducted on two multi-labeled cross-modality cardiac datasets and one abdominal multi-organ dataset, which demonstrate that our method outperforms other state-of-the-art methods.