[Paper Review] Deep Symmetric Adaptation Network for Cross-modality Medical Image Segmentation

Xiaoting Han, Lei Qi+5 more|arXiv (Cornell University)|Jan 18, 2021

Domain Adaptation and Few-Shot Learning36 references4 citations

TL;DR

This paper proposes a deep symmetric adaptation network for cross-modality medical image segmentation, using bidirectional feature alignment via shared encoders and private decoders in two symmetric translation sub-networks. The method improves segmentation performance by jointly training on real and translated images from both domains, achieving state-of-the-art results on Cardiac and BraTS datasets with a Dice score of 78.50%.

ABSTRACT

Unsupervised domain adaptation (UDA) methods have shown their promising performance in the cross-modality medical image segmentation tasks. These typical methods usually utilize a translation network to transform images from the source domain to target domain or train the pixel-level classifier merely using translated source images and original target images. However, when there exists a large domain shift between source and target domains, we argue that this asymmetric structure could not fully eliminate the domain gap. In this paper, we present a novel deep symmetric architecture of UDA for medical image segmentation, which consists of a segmentation sub-network, and two symmetric source and target domain translation sub-networks. To be specific, based on two translation sub-networks, we introduce a bidirectional alignment scheme via a shared encoder and private decoders to simultaneously align features 1) from source to target domain and 2) from target to source domain, which helps effectively mitigate the discrepancy between domains. Furthermore, for the segmentation sub-network, we train a pixel-level classifier using not only original target images and translated source images, but also original source images and translated target images, which helps sufficiently leverage the semantic information from the images with different styles. Extensive experiments demonstrate that our method has remarkable advantages compared to the state-of-the-art methods in both cross-modality Cardiac and BraTS segmentation tasks.

Motivation & Objective

• Address the large domain shift between different medical imaging modalities (e.g., MRI and CT) that degrades segmentation performance.
• Overcome limitations of asymmetric UDA methods that rely on unidirectional translation or feature alignment.
• Improve segmentation generalization by symmetrically aligning features across source and target domains.
• Leverage both real and translated images from both domains to enhance semantic feature learning.
• Achieve state-of-the-art performance in cross-modality medical image segmentation tasks.

Proposed method

• Introduce a symmetric architecture with two translation sub-networks: one translating source to target and another translating target to source, both sharing the same encoder.
• Use shared encoders to learn domain-invariant features and private decoders to learn domain-specific representations via reconstruction losses.
• Apply adversarial training with shared discriminators to distinguish real vs. generated images in both domains, ensuring semantic consistency.
• Train the segmentation network using real target images, translated source images, real source images, and translated target images to fully exploit multi-style semantic information.
• Implement bidirectional feature alignment by jointly optimizing the segmentation and translation sub-networks in an end-to-end manner.
• Utilize a shared discriminator for both domains to maintain consistency in adversarial loss and improve feature generalization.

Experimental results

Research questions

• RQ1 Can symmetric domain adaptation via bidirectional translation improve cross-modality medical image segmentation compared to asymmetric approaches?
• RQ2 Does joint training on real and translated images from both source and target domains enhance semantic feature learning and segmentation accuracy?
• RQ3 How does sharing the encoder across segmentation and translation networks affect domain-invariant representation learning?
• RQ4 What is the impact of using private decoders versus shared decoders in the translation sub-networks?
• RQ5 How effective is the use of shared discriminators in maintaining semantic consistency across real and generated images?

Key findings

• The proposed method achieves a Dice score of 78.50% on the BraTS dataset, outperforming state-of-the-art methods.
• The method reduces the average surface distance (ASD) to 5.92, indicating improved segmentation boundary accuracy.
• Using private decoders instead of shared decoders improves Dice by 33.57 percentage points (from 44.93% to 78.50%).
• Sharing the encoder improves Dice by 4.60 percentage points compared to using unshared encoders.
• Sharing discriminators boosts Dice by 2.36 percentage points compared to using unshared discriminators.
• The model achieves significant performance gains on both Cardiac and BraTS datasets, demonstrating robustness across different medical imaging tasks.