A cloud-edge adaptive lightweight network with dynamic inference enables real-time ultrasound image segmentation.

Ultrasound image segmentation plays an important role in clinical diagnosis, but existing deep learning methods struggle to balance accuracy and efficiency, limiting practical deployment in resource-constrained medical scenarios. This study proposes a Cloud-Edge Adaptive Lightweight Network (CEA-Net) with dynamic inference, implementing parameter compression through a lightweight encoder constructed with depthwise separable convolutions and inverted residual structures, introducing serial channel-spatial attention mechanisms to enhance lesion feature representation, and designing intelligent routing algorithms based on sample complexity perception to dynamically select edge fast inference or cloud enhanced processing paths according to ultrasound image complexity. Experimental validation on the BUSI breast ultrasound dataset and TN3K thyroid nodule dataset demonstrates that CEA-Net significantly reduces model parameters and inference latency while maintaining high segmentation accuracy, achieves near real-time inference capability on edge devices, and reaches a favourable balance between accuracy and speed through intelligent task allocation in cloud-edge collaborative mode. This study provides a feasible technical solution for practical deployment of ultrasound image segmentation technology in resource-constrained scenarios including primary healthcare institutions, telemedicine, and mobile diagnosis, advancing the practical application of medical artificial intelligence technology from cloud to edge.