Dual Co-Train: Cross-Dataset Ultrasound Tongue Segmentation Under Extreme Data Scarcity

Alisher Myrgyyassov; Zhen Song; Bruce Xiao Wang; Yu Sun; Min Ney Wong; Yihao Zhou; Yongping Zheng

Source-linked AI summary

Dual Co-Train: Cross-Dataset Ultrasound Tongue Segmentation Under Extreme Data Scarcity

Alisher Myrgyyassov, Zhen Song, Bruce Xiao Wang, Yu Sun, Min Ney Wong, Yihao Zhou, Yongping Zheng

arXiv:2608.17983v1Aug 18, 2026cs.CVcs.AI

TL;DR

Cross-dataset ultrasound tongue segmentation remains difficult under domain shift and limited annotation. This source-free dual co-training framework refines pseudo-labels and uses target-style synthetic augmentation, achieving strong adaptation especially when labeled source data are scarce.

Problem
Cross-dataset ultrasound tongue segmentation lacks broad evaluation and practical methods that reduce annotation demands under varied imaging conditions.
Method
A source-free dual co-training framework jointly refines target pseudo-labels and synthesizes target-style image–mask pairs for segmentation adaptation.
Results
0.713 Dice with 5 labeled images matched the supervised baseline’s 0.716 Dice using 1,611 images in MTID(Legacy)→MTID.
Takeaways & Limitations
The framework is especially effective for cross-dataset and in-dataset adaptation when labeled source data are limited.
Takeaways & Limitations
Future work is needed to test additional ultrasound corpora and reduce reliance on synthetic generation while preserving robustness under severe domain shift.

Abstract

from arXiv · show

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.

1 Introduction

This work frames ultrasound tongue contour segmentation as a low-label, source-free cross-dataset adaptation problem under substantial imaging variability. It adapts a five-image-pretrained UltraUNet using unlabeled target frames through pseudo-label refinement, contour quality control, synthetic augmentation, and co-training.

UTI contour segmentation is challenged by probe placement, acquisition settings, operator skill, anatomy, and motion artifacts that induce domain variability.
The framework initializes UltraUNet with only five labeled source images and adapts it to a new target dataset using unlabeled target images.
The study evaluates transfer across eight heterogeneous UTI datasets spanning different speakers, recording setups, transducers, and imaging artifacts.
Validation includes cross-dataset experiments, supervised and source-free baseline comparisons, and ablations assessing each design choice's contribution to generalization.
The proposed adaptation combines pseudo-labeling, contour-based quality control, conditional GAN synthetic augmentation, and GAN-segmenter co-training.

2 Related Works

Prior ultrasound tongue imaging research progressed from classical contour methods to deep segmentation, but evaluations and transfer-learning studies remain limited under substantial cross-dataset shifts. SSL and SFDA offer relevant tools, yet existing UTI work does not directly address source-free contour adaptation, motivating the proposed framework.

Ultrasound tongue imaging: UTI research evolved from active contours, active appearance models, and Markov random fields to CNN and encoder-decoder methods for contour tracking and segmentation.
Cross-dataset generalization: Most UTI segmentation studies remain within single datasets or recording environments, limiting assessment of generalization across imaging conditions, subjects, devices, and language backgrounds.
Cross-dataset generalization: Transfer learning reuses features across speakers or corpora but typically evaluates modest domain shifts, leaving large hardware, protocol, and population differences insufficiently addressed.
Related architectures: UltraUNet provides a lightweight, high-framerate UTI segmentation architecture and a multi-dataset benchmark spanning multiple corpora and acquisition conditions.
Semi-supervised and source-free adaptation: SSL methods exploit unlabeled images through consistency regularization, pseudo-labeling, cross-teaching, and teacher-student frameworks, but usually assume the same or closely related domain.
Semi-supervised and source-free adaptation: UTI unlabeled-data studies mainly use self-supervised, reconstruction-based, or cross-modal learning, rather than source-free contour segmentation and adaptation under cross-dataset shift.This gap motivates the dual co-training framework proposed in this work.

3 Methods

The method performs source-free cross-dataset adaptation by jointly refining target pseudo-labels and synthesizing target-style image–mask pairs. It uses UltraUNet, contour-based quality control, student–teacher consistency training, and periodic GAN updates without target annotations.

Dual Co-Train Framework: The framework couples an UltraUNet segmenter with a segmentation-guided conditional GAN for iterative pseudo-label refinement and target-style image–mask synthesis.It adapts a source-trained contour segmenter to an unlabeled target domain while improving segmentation robustness through dual co-training.
Training Stages: Adaptation proceeds through source pretraining, GAN pretraining on generated pseudo-masks, and target fine-tuning with pseudo-label-driven dual co-training.The target domain is held out during testing, and no target labels are used during adaptation.
Pseudo-Label Quality Control: A probability threshold of 0.5 and contour-based quality control divide teacher-generated target pseudo-masks into clean and noisy subsets.The rule-based QC module rejects tiny components, fragmented or broken contours, border-touching masks, excessive holes, and unstable anatomical shapes.
Dual Co-Training Objective: The student trains on clean pseudo-labels with supervised loss, noisy pseudo-labels with Mean Teacher consistency, and GAN-generated target-style pairs with supervised loss.The teacher is maintained as an exponential moving average of the student, with λ = 0.99 in all experiments.
Closed-Loop GAN Updating: Every four segmentation epochs, the teacher re-infers target pseudo-masks and the GAN is fine-tuned for two epochs from the current best checkpoint.This periodic refresh keeps the synthetic image distribution aligned with the evolving target pseudo-mask distribution.

4 Experiments and Results

Across 12 source-to-target transfer pairs, Dual Co-Train achieved the best average Dice and MSD, with statistically significant gains over all five source-free baselines. Ablations and scaling studies show that pseudo-label quality control is especially important and that the method reaches strong performance with far fewer labeled source images, while within-dataset transfers remain easier than cross-dataset transfers.

Overall comparison: The method achieved the highest average Dice (0.760) and lowest MSD (2.412 px), outperforming EMA, SHOT, UPL, AIF, and FSM across the 12 transfer pairs.It attained the best Dice in 10 of 12 pairs.
Statistical analysis: Friedman tests found significant differences for Dice (χ2 = 38.81, df = 5, p < 0.001) and MSD (χ2 = 32.10, df = 5, p < 0.001), while corrected Wilcoxon tests confirmed superiority over all five baselines.The post-hoc tests used Bonferroni correction with α = 0.01.
Ablation study: Removing the EMA teacher, GAN refinement, or pseudo-label quality control degraded both metrics, with the QC ablation producing the largest Dice reduction to 0.735 on average.The full model achieved 2.412 px MSD and 0.760 Dice on average.
Source-size scaling: For MTID(Legacy)→MTID, Dual Co-Train reached 0.713 Dice with N = 5, matching the baseline’s 0.716 Dice using all 1,611 images.For UXTD(Legacy)→MTID, it achieved 0.639 Dice at N = 1, compared with the baseline peak of 0.640 Dice at N = 200.
Source-size scaling: Dual Co-Train exhibited visibly narrower variance bands across seeds at every source training size, indicating lower sensitivity to which images were sampled.The baseline showed wider and less stable confidence intervals, particularly at small N.
Transfer difficulty: Within-dataset transfers reached higher final Dice scores for MTID(Legacy)→MTID (0.755) and UXTD(Legacy)→UXTD (0.778) than cross-dataset transfers at 0.726 and 0.699.The reported results are consistent with less domain shift in same-dataset transfer.

5 Discussion

The discussion concludes that dual co-training is especially effective for low-label, source-free ultrasound tongue segmentation under strong domain shift, while pseudo-label quality control and synthetic-branch refinement are central to its gains. Scaling results indicate substantial labeling-efficiency benefits, although severe appearance mismatch and external-method comparisons remain limitations.

Key findings: The framework is particularly effective when labeled target and source data are unavailable, with its largest improvements appearing in the low-data regime.The discussion attributes this pattern to substantial domain shift and scaling-experiment results.
Cross-domain behavior: Transfers from cleaner MTID and CTID to noisy TaL1 and Cleft often performed better than expected, whereas noisy TaL1→UXTD remained challenging.The discussion attributes the contrast to domain-specific augmentation, severe appearance mismatch, and annotation noise.
Scaling and efficiency: 0.713 Dice with 5 labeled images matched the supervised baseline’s 0.716 Dice using 1,611 images in MTID(Legacy)→MTID.This comparison is presented as evidence of labeling-efficiency gains rather than transfer improvement alone.
Ablation findings: Removing pseudo-label quality control caused the largest Dice drop, while removing periodic GAN refinement also consistently degraded performance.Quality control limits noisy-mask propagation, and GAN refreshing keeps the synthetic image distribution aligned with evolving target masks.
Limitations: Direct comparison with external source-free adaptation methods is difficult because their backbones, losses, schedules, and pretrained-model assumptions differ from the low-label UltraUNet setting.Faithful reimplementation would require substantial redesign and additional hyperparameter tuning, making comparisons less controlled.
Scaling and efficiency: Comparable Dice was often reached within 10-50 labeled images rather than several hundred, substantially reducing source annotation requirements.The passage links this reduction to pseudo-label refinement and synthetic augmentation recovering most of a dataset’s achievable performance ceiling.

6 Conclusion

The paper concludes that its source-free dual co-training framework improves cross-dataset ultrasound tongue segmentation under low-label conditions while remaining practical for real-time use. It emphasizes task-specific adaptation and identifies uncertainty modeling, broader corpora, and more efficient adaptation as future directions.

Conclusion: The source-free dual co-training framework combines pseudo-label self-training, contour-based quality control, and segmentation-guided synthetic augmentation within lightweight UltraUNet.The framework targets cross-dataset ultrasound tongue contour segmentation under low-label conditions.
Conclusion: Experiments across 12 transfer pairs, scaling studies, and ablations support improved adaptation performance while retaining practical real-time use.The supplied conclusion passage reports these evaluation settings but is truncated before stating detailed quantitative results.
Conclusion: Task-specific adaptation addresses domain shift, annotation variability, and limited supervision by using structural contour information and target-domain synthesis.The approach is presented as a promising option for reproducible and label-efficient tongue segmentation across transfer scenarios.
Future work: Future work will extend evaluation to additional ultrasound corpora, strengthen uncertainty modeling for pseudo-label selection, and improve adaptation efficiency.One stated goal is reducing reliance on synthetic generation while preserving robustness under severe domain shift.