Contrastive adversarial mix-up learning for unsupervised cross-modality image segmentation

Background: Unsupervised domain adaptation (UDA) has been extensively investigated in terms of its ability to address the performance degradation caused by distribution shifts between labeled source-domain images and unlabeled target-domain images. In radiotherapy, this problem is particularly relevant for the transfer of segmentation knowledge from planning computed tomography (CT) to cone-beam computed tomography (CBCT), for which target-domain annotations are limited. This study aimed to develop a contrastive adversarial mix-up learning framework for unsupervised cross-modality image segmentation.

Methods: A UDA framework integrating mix-up augmentation (MA) and contrastive adversarial learning (AL) was developed. This framework included a shared-weight redistribution network to reduce cross-domain appearance discrepancies, a mix-up consistency strategy to constrain segmentation predictions of mixed-domain images, and contrastive AL to regulate the relationship between redistributed mixture images and mixed redistributed images. The method was evaluated with two in-house CT-CBCT datasets comprising 112 patients with nasopharyngeal carcinoma and 112 patients with breast cancer, with 12 patients from each dataset randomly selected for testing. Generalizability was further evaluated in 75 patients from the public Brain Tumor Segmentation 2018 (BraTS2018) dataset, with 80% used for training and 20% for testing. Segmentation performance was assessed via the Dice similarity coefficient (DSC), Hausdorff distance (HD), and Jaccard index.

Results: On the in-house CT-CBCT datasets, the complete UDA + MA + AL framework achieved DSC values of 82.08%, 85.81%, 78.60%, and 60.75% for the breast cancer clinical target volume (CTV), heart, head-and-neck cancer CTV, and parotid glands, respectively; meanwhile, compared with the UDA-only framework, the corresponding absolute DSC improvements of the complete framework were 8.70, 1.02, 3.28, and 3.04 percentage points, respectively. On the BraTS2018 dataset, the proposed framework achieved DSC values of 63.56%, 67.81%, 85.47%, and 74.16% for tasks adapting T2-weighted magnetic resonance image (T2) to T1-weighted magnetic resonance image (T1), T2 to T1 contrast enhanced, T2 to fluid-attenuated inversion recovery (FLAIR), and FLAIR to T2, respectively. Overall, the proposed method achieved an average DSC improvement of 3.53 percentage points compared with representative UDA-based segmentation methods.

Conclusions: The proposed contrastive adversarial mix-up learning framework improved unsupervised cross-modality segmentation performance across CT-CBCT and multisequence magnetic resonance imaging adaptation tasks. By combining cross-domain redistribution, mix-up consistency learning, and contrastive AL, the framework can provide enhance robustness for variations in image appearance and anatomical structure while maintaining segmentation consistency.