Dual Co-Train: New AI Framework for Ultrasound Tongue Segmentation Under Extreme Data Scarcity
A paper titled 'Dual Co-Train: Cross-Dataset Ultrasound Tongue Segmentation Under Extreme Data Scarcity' has been released on arXiv. It presents a framework for source-free domain adaptation aimed at overcoming cross-dataset domain shifts in ultrasound tongue contour segmentation, particularly in scenarios with few annotations and noise during acquisition. Utilizing a lightweight UltraUNet backbone, the method begins with a checkpoint that is pretrained on five labeled images, mimicking an underfitted model. It employs a closed-loop process to adapt to an unlabeled target domain, enhancing pseudo-labels and creating synthetic image-mask pairs through a segmentation-guided GAN. This strategy proves effective for adaptation in clinical environments where annotated ultrasound data is limited. The arXiv identifier for the paper is 2608.17983.
Key facts
- The paper is titled 'Dual Co-Train: Cross-Dataset Ultrasound Tongue Segmentation Under Extreme Data Scarcity'.
- It is available on arXiv under identifier 2608.17983.
- The framework is source-free, meaning it does not require access to source domain data during adaptation.
- It uses a lightweight UltraUNet backbone.
- The source model is pretrained on only five labeled source images.
- The target domain is fully unlabeled.
- The method includes a contour-based quality-control module to filter unreliable masks.
- It employs a segmentation-guided conditional GAN to generate target-style synthetic image-mask pairs.
- Training combines clean pseudo-labels, noisy pseudo-labels with consistency regularization, and synthetic samples.
Entities
Institutions
- arXiv