Enabling Real-Time Point-of-Care Ultrasound Segmentation: A GPU-Free Deployment in Resource-Limited Settings

The article presents UltraSeg, a lightweight architecture adapted for real-time point-of-care ultrasound (POCUS) segmentation, with two variants: UltraSeg-130K (0.13M parameters) and UltraSeg-500K (0.5M parameters). UltraSeg-130K achieves 89.7 FPS on single-core CPUs, while UltraSeg-500K matches or exceeds the performance of larger models like UNet and TransUNet, demonstrating superior cross-dataset generalization. This development facilitates AI deployment in resource-limited settings by eliminating the need for GPU acceleration, thus making advanced diagnostic capabilities more accessible.