Researchers have developed a method to improve the accuracy of ultrasound tongue segmentation under extreme data scarcity. The approach, called Dual Co-Train, uses pseudo-label refinement and synthetic target-style augmentation to adapt to new datasets without access to source data. In experiments involving eight ultrasound tongue imaging datasets, the proposed framework outperformed baseline methods, including supervised ones, in terms of segmentation overlap and contour accuracy. --- Why it matters: This matters because accurate ultrasound tongue segmentation is crucial for speech therapy and other medical applications, but collecting labeled training data can be challenging due to variability in probe settings and acquisition noise. Dual Co-Train's ability to adapt to new datasets without source data could make it a valuable tool for researchers and clinicians. Source: https://arxiv.org/abs/2608.17983