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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning for ultrasound image segmentation faces challenges in balancing accuracy and efficiency.
• Practical deployment is limited in resource-constrained medical environments.

Purpose of the Study:

• To propose a Cloud-Edge Adaptive Lightweight Network (CEA-Net) for efficient and accurate ultrasound image segmentation.
• To address the limitations of current methods in real-world clinical settings.

Main Methods:

• Developed CEA-Net with a lightweight encoder using depthwise separable convolutions and inverted residual structures.
• Incorporated serial channel-spatial attention mechanisms for enhanced lesion feature representation.
• Designed intelligent routing algorithms for dynamic selection between edge fast inference and cloud enhanced processing based on image complexity.

Main Results:

• CEA-Net significantly reduced model parameters and inference latency on BUSI and TN3K datasets.
• Achieved high segmentation accuracy, approaching near real-time inference on edge devices.
• Demonstrated a favorable balance between accuracy and speed via intelligent cloud-edge task allocation.

Conclusions:

• CEA-Net offers a feasible solution for deploying ultrasound image segmentation in resource-limited scenarios like primary care and mobile diagnostics.
• The study advances practical medical AI applications by enabling cloud-to-edge collaborative processing.
• This technology supports improved clinical diagnosis and telemedicine capabilities.