Cloud-to-Edge Deployment of Optimized nnU-Net for Ischemic Stroke Lesion Segmentation on Resource-Constrained...

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Neurology

Background:

• Ischemic stroke necessitates rapid, accurate lesion assessment for effective treatment.
• The "Time is Brain" principle highlights the urgency in stroke care.
• Advanced AI models like nnU-Net v2 offer state-of-the-art medical image segmentation but face deployment challenges in resource-constrained settings.

Purpose of the Study:

• To assess the technical feasibility of deploying a 3D nnU-Net v2 model from the cloud to an embedded edge device.
• To evaluate the performance and efficiency of a Cloud-to-Edge optimization pipeline for medical image segmentation.

Main Methods:

• A 3D nnU-Net v2 model was optimized for deployment on a resource-constrained embedded device using a Cloud-to-Edge pipeline.
• Segmentation performance was evaluated using overlap-based metrics (e.g., Dice score).
• Inference speed and power consumption were measured for different precision configurations (FP32, FP16) using TensorRT.

Main Results:

• Edge deployment resulted in a decrease in segmentation metrics (Dice from ~0.78 to 0.67) compared to cloud-based inference.
• TensorRT FP32 and FP16 inference showed comparable segmentation performance, indicating minimal degradation with reduced precision.
• The optimized FP16 configuration achieved a processing time of 10.2 seconds per 3D volume (33% faster than FP32) at low power consumption (10-13 W).

Conclusions:

• The study demonstrates the preliminary technical feasibility of running advanced 3D volumetric segmentation models on low-power edge hardware.
• Optimized edge deployment offers a potential pathway for real-time stroke imaging analysis at the point-of-care.
• Further validation, including external datasets and clinical assessment, is required before clinical application.