Domain-Agnostic Learning With Anatomy-Consistent Embedding for Cross-Modality Liver Segmentation

Abstract

Domain Adaptation (DA) has the potential to greatly help the generalization of deep learning models. However, the current literature usually assumes to transfer the knowledge from the source domain to a specific known target domain. Domain Agnostic Learning (DAL) proposes a new task of transferring knowledge from the source domain to data from multiple heterogeneous target domains. In this work, we propose the Domain-Agnostic Learning framework with Anatomy-Consistent Embedding (DALACE) that works on both domain-transfer and task-transfer to learn a disentangled representation, aiming to not only be invariant to different modalities but also preserve anatomical structures for the DA and DAL tasks in cross-modality liver segmentation. We validated and compared our model with state-of-the-art methods, including CycleGAN, Task Driven Generative Adversarial Network (TD-GAN), and Domain Adaptation via Disentangled Representations (DADR). For the DA task, our DALACE model outperformed CycleGAN, TD-GAN, and DADR with DSC of 0.847 compared to 0.721, 0.793 and 0.806. For the DAL task, our model improved the performance with DSC of 0.794 from 0.522, 0.719 and 0.742 by CycleGAN, TD-GAN, and DADR. Further, we visualized the success of disentanglement, which added human interpretability of the learned meaningful representations. Through ablation analysis, we specifically showed the concrete benefits of disentanglement for downstream tasks and the role of supervision for better disentangled representation with segmentation consistency to be invariant to domains with the proposed Domain-Agnostic Module (DAM) and to preserve anatomical information with the proposed Anatomy-Preserving Module (APM).

OLVA: Optimal Latent Vector Alignment for Unsupervised Domain Adaptation in Medical Image Segmentation

“…few labeled instances in the target domain can be used for joint-training with the source data [19]. The more ambitious UDA strategies [1,3,4,17,23,24] assume no labels are available for the target domain. The core idea of UDA is to go through an adaption phase using a non-linear mapping to find a common domaininvariant representation or a latent space Z.…”

“…Related Work. Recent works on UDA for medical image segmentation rely on Generative Adversarial Networks [3,4,7,17,23,24] to translate the appearance from one modality to the other using multiple discriminators and a pixelwise cycle consistency loss. Despite their success, they: i) suffer from instabilities [2], ii) rely on complex architectures with more than 95 million parameters, iii) are prone to model collapse [16], and iv) may generate images outside the actual target domain [1].…”

Optimal Latent Vector Alignment for Unsupervised Domain Adaptation in Medical Image Segmentation

“…labeled instances in the target domain can be used for joint-training with the source data [19]. The more ambitious UDA strategies [1,3,4,17,23,24] assume no labels are available for the target domain. The core idea of UDA is to go through an adaption phase using a non-linear mapping to find a common domaininvariant representation or a latent space Z.…”

“…Related Work. Recent works on UDA for medical image segmentation rely on Generative Adversarial Networks [3,4,7,17,23,24] to translate the appearance from one modality to the other using multiple discriminators and a pixel-wise cycle consistency loss. Despite their success, they: i) suffer from instabilities [2], ii) rely on complex architectures with more than 95 million parameters, iii) are prone to model collapse [16], and iv) may generate images outside the actual target domain [1].…”

Partly Supervised Multitask Learning

“…In unsupervised domain adaptation, a model is pre-trained on similar tasks in some other domains with labeled data, and the pre-trained model is then fine-tuned with a limited set of labeled data in the target domain. Domain adaption can also be performed to learn a generic representation where the model is fully-supervised for source data and unsupervised for the target data (Yang et al, 2019;Zhuang et al, 2019). Self-supervised learning is closely related to transfer learning.…”