Search papers, labs, and topics across Lattice.
This paper introduces a source-free domain adaptation framework for ultrasound tongue contour segmentation that operates under extreme data scarcity, using a lightweight UltraUNet backbone. By starting with a model pretrained on just five labeled images, the method iteratively refines pseudo-labels and employs a contour-based quality-control module to enhance segmentation accuracy. The results demonstrate significant improvements in segmentation overlap and contour accuracy across multiple datasets, highlighting the effectiveness of pseudo-label refinement and synthetic augmentation in challenging cross-dataset scenarios.
Task-specific pseudo-label refinement and synthetic augmentation can dramatically enhance segmentation performance even with minimal labeled data.
Ultrasound tongue contour segmentation remains challenging under cross-dataset domain shift, where limited annotations, probe variability, and acquisition noise often degrade model generalization. We present a source-free domain adaptation framework for robust ultrasound tongue segmentation built on a lightweight UltraUNet backbone. Starting from a checkpoint pretrained on only five labeled source images, simulating an underfitted constrained source model, the proposed method adapts to a fully-unlabeled target domain by iteratively refining pseudo-labels, filtering unreliable masks with a contour-based quality-control module, and generating target-style synthetic image-mask pairs through a segmentation-guided conditional GAN. The student model is then trained on a mixture of clean pseudo-labeled target images, noisy pseudo-labels with consistency regularization, and synthetic samples, enabling closed-loop adaptation without access to source data. We evaluate the method on 12 source-target transfer pairs across eight ultrasound tongue imaging datasets, and conduct source-size scaling experiments and ablation studies. Across all comparisons, the proposed framework improves segmentation overlap and contour accuracy over the baselines, including supervised ones. These results suggest that task-specific pseudo-label refinement and synthetic target-style augmentation can substantially improve source-free adaptation for ultrasound tongue imaging.
