Related Experiment Video
Updated: Aug 13, 2026
High Frequency Ultrasound for the Analysis of Fetal and Placental Development In Vivo
Published on: November 8, 2018
Text-prompted foundation-model segmentation of intrapartum fetal ultrasound: a controlled methodological study with
Zhiyuan Yang1, Qinghua Meng1, Bentao Jiang2
1Tianjin University of Sport, Tianjin, China.
Background:
Segmenting intrapartum fetal anatomy on B-mode ultrasound supports labor-progress monitoring, but conventional workflows require task-specific models and manual geometric prompts. We investigate whether text-conditioned prompting can select anatomical targets under controlled-label budgets.
Methods:
We propose F-TPA-SAM2, a Text-to-Prompt Adapter with 0.871 M trainable parameters (0.135% of the model), which maps frozen FetalCLIP embeddings to a low-rank dense prompt for the frozen SAM 2 mask decoder. We trained the adapter on PSFHS using three seeds and label budgets N∈{20, 100, 400}. Controls included no-text and sparse-only variants, blank and wrong-class prompts, a parameter-matched no-text adapter, and fixed random class codes. A supervised U-Net was trained from scratch under the same N = 100 and N = 400 budgets. Comparisons used image-clustered paired bootstrap confidence intervals and permutation tests. We also assessed zero-shot transfer to 779 HC18 images and compared FetalCLIP, BiomedCLIP, and OpenCLIP at N = 100.
Results:
On the PSFHS test set, F-TPA-SAM2 achieved 0.811 ± 0.011 Dice at N = 100 and 0.859 ± 0.004 at N = 400. Differences from the parameter-matched no-text adapter and the fixed random class codes were unclear at both budgets because all paired 95% confidence intervals included zero. F-TPA-SAM2 exceeded U-Net by 0.053 Dice at N = 100 (95% CI 0.006-0.085), whereas the N = 400 difference was unclear. Wrong-class prompts redirected predictions to the opposite class: requested-class Dice was 0.0003, whereas opposite-class Dice was 0.859 ± 0.003. Zero-shot HC18 Dice was 0.812 ± 0.157, and the three text encoders were within 0.009 Dice.
Conclusion:
F-TPA-SAM2 uses a stable category-conditioned input to select between the two trained anatomical targets. However, matched no-text and random-code controls indicate that pre-trained language semantics are not necessary for this result. The findings do not establish open-vocabulary or multi-anatomy segmentation; clinical utility requires a sonographer-in-the-loop and prospective validation.
