Ultra-ECP: Ellipse-Constrained and Point-Robust Foundation Model Adaptation for Fetal Cardiac Ultrasound Segmentation

Abstract

Accurate fetal cardiac segmentation from four-chamber ultrasound images is essential for reliable prenatal biometrics, yet foundation models such as SAM remain sensitive to point-prompt placement, produce anatomically inconsistent masks, and require costly full-model fine-tuning. We introduce Ultra-ECP, a parameter-efficient framework that adapts UltraSAM for robust single-point fetal cardiac segmentation. Ultra-ECP integrates three components: (i) a LoRA-based adaptation applied to the prompt encoder and mask decoder, reducing trainable parameters by over 98%; (ii) an Ellipse-Aware Loss that regularizes predictions toward anatomically plausible elliptical cardiac shapes; and (iii) a Point-Robust Augmentation strategy that simulates click imprecision to enhance robustness. Evaluated on the FOCUS dataset, Ultra-ECP outperforms SAM, MedSAM, and fine-tuned U-Net baselines. For thoracic segmentation, it achieves a mean DSC of 95.09% and HD95 of 25.96 px. For cardiac segmentation, Ultra-ECP obtains a mean DSC of 92.60% and HD95 of 18.25 px, while maintaining stability under point displacements of up to 10 pixels. Predictions are consistently smooth and elliptical, addressing common failure modes of existing approaches. Ultra-ECP provides an effective and computationally lightweight pathway for adapting large vision models to fetal cardiac biometrics, enabling reliable and clinically practical semi-automated tools.