[Paper Review] Cross-Modality Brain Tumor Segmentation via Bidirectional Global-to-Local Unsupervised Domain Adaptation

Kelei He, Wen Ji+8 more|arXiv (Cornell University)|May 17, 2021

Domain Adaptation and Few-Shot Learning23 references4 citations

TL;DR

This paper proposes a bidirectional global-to-local unsupervised domain adaptation framework (BiGL) for cross-modality brain tumor segmentation using multi-modal MRI. By jointly synthesizing images in both directions and enforcing global-to-local consistency in feature representations, BiGL significantly improves segmentation accuracy across T1 and T2 MRI sequences, outperforming state-of-the-art methods by large margins on benchmark datasets.

ABSTRACT

Accurate segmentation of brain tumors from multi-modal Magnetic Resonance (MR) images is essential in brain tumor diagnosis and treatment. However, due to the existence of domain shifts among different modalities, the performance of networks decreases dramatically when training on one modality and performing on another, e.g., train on T1 image while performing on T2 image, which is often required in clinical applications. This also prohibits a network from being trained on labeled data and then transferred to unlabeled data from a different domain. To overcome this, unsupervised domain adaptation (UDA) methods provide effective solutions to alleviate the domain shift between labeled source data and unlabeled target data. In this paper, we propose a novel Bidirectional Global-to-Local (BiGL) adaptation framework under a UDA scheme. Specifically, a bidirectional image synthesis and segmentation module is proposed to segment the brain tumor using the intermediate data distributions generated for the two domains, which includes an image-to-image translator and a shared-weighted segmentation network. Further, a global-to-local consistency learning module is proposed to build robust representation alignments in an integrated way. Extensive experiments on a multi-modal brain MR benchmark dataset demonstrate that the proposed method outperforms several state-of-the-art unsupervised domain adaptation methods by a large margin, while a comprehensive ablation study validates the effectiveness of each key component. The implementation code of our method will be released at \url{https://github.com/KeleiHe/BiGL}.

Motivation & Objective

• To address the challenge of domain shift between different MRI modalities (e.g., T1 to T2) that degrades segmentation performance when models are trained on one modality and tested on another.
• To overcome limitations in existing unsupervised domain adaptation (UDA) methods, such as unidirectional adaptation, inefficient feature alignment, and poor handling of tumor-specific structural variability.
• To develop a flexible and generalizable framework that enables high-accuracy segmentation on unlabeled target modality data using only labeled source modality data.
• To improve model generalization by integrating bidirectional image synthesis and multi-scale, attention-based representation alignment.

Proposed method

• Introduces a bidirectional image-to-image translation module that generates synthesized images for both source and target domains to bridge domain gaps.
• Employs a shared-weight encoder-decoder segmentation network to predict tumor masks on original and synthesized images, enabling joint optimization.
• Applies a global-to-local consistency learning module that aligns features across domains at both global and local levels using attention mechanisms.
• Uses adversarial training in the image translation branch to ensure realistic synthesis of modality-specific images.
• Incorporates attention-guided feature alignment to focus on discriminative tumor regions and suppress irrelevant background features.
• Enforces consistency between original and synthesized images at multiple levels to improve robustness and generalization.

Experimental results

Research questions

• RQ1 Can bidirectional image synthesis between MRI modalities improve cross-modality tumor segmentation performance under unsupervised domain adaptation?
• RQ2 How does global-to-local feature alignment enhance representation consistency and segmentation accuracy across domains with large domain shifts?
• RQ3 Does incorporating attention mechanisms into domain adaptation improve the model's ability to focus on tumor-relevant regions during adaptation?
• RQ4 How does the proposed method compare to state-of-the-art UDA methods in terms of segmentation performance on brain tumor and cardiac substructure datasets?
• RQ5 Can the framework generalize to other MRI sequences such as T1 and FLAIR without retraining?

Key findings

• On the BraTS 2018 dataset, BiGL achieved a mean DSC of 10.9% on T1-to-T2 adaptation and 5.4% on T2-to-T1, outperforming all baselines including SIFA-TMI and Cycada.
• In cross-modality cardiac substructure segmentation on MMWHS, BiGL achieved a DSC of 75.3% and ASD of 6.2 for MR-to-CT adaptation, surpassing SIFA-TMI by 6.4% in DSC for the AA class.
• BiGL reduced the ASD by 1.5–2.0 mm compared to SIFA-TMI and AdaptSegNet, indicating more precise contour predictions.
• Ablation studies confirmed that both bidirectional synthesis and global-to-local consistency learning are essential for performance gains.
• The method demonstrated robustness across diverse modality pairs, with visualizations showing significantly improved segmentation contours close to ground truth.
• The implementation code is publicly released at https://github.com/KeleiHe/BiGL, enabling reproducibility and extension to other medical imaging tasks.