A 23-µJ-per-Frame All-on-Chip TinyML U-Net Processor for Real-Time Autonomous Image Segmentation in Miniaturized...

Area of Science:

• Biomedical Engineering
• Computer Engineering
• Artificial Intelligence

Background:

• Autonomous medical image segmentation is crucial for wearable ultrasound patches, requiring on-device processing for privacy and functionality.
• Existing binary neural network (BNN) accelerators lack features for medical-grade segmentation, facing accuracy, utilization, and memory challenges.
• U-Net architectures offer high performance, but their computational demands are challenging for resource-constrained wearable devices.

Purpose of the Study:

• To develop a fully-integrated U-Net processor for efficient, on-device medical image segmentation in wearable ultrasound patches.
• To overcome the limitations of existing BNN accelerators in supporting accurate and computationally efficient medical image segmentation.
• To enable real-time, autonomous monitoring capabilities in wearable medical devices through specialized hardware.

Main Methods:

• Designed a 0.81 mm² integrated processor in 28nm technology featuring mixed-precision datapaths (binary convolution with 4-bit skip connections).
• Conducted systematic design space exploration across 9,390 configurations to optimize energy-latency tradeoffs.
• Implemented interleaved memory representation, halo reuse, hardware-supported layer fusion, and lossless compression to eliminate external memory dependency.

Main Results:

• The processor achieves 13.4 frames per second (fps) and 23 µJ per frame on bladder and fetal head segmentation datasets.
• Demonstrated significant reductions in peak on-chip memory usage by 3.16× (layer fusion) and 1.38× (lossless compression).
• Validated the processor's capability for clinical accuracy and energy-efficient, battery-powered operation in wearable medical devices.

Conclusions:

• The developed integrated U-Net processor enables accurate and efficient on-device medical image segmentation for wearable ultrasound applications.
• The mixed-precision design and memory optimization techniques address key limitations of prior BNN accelerators for medical use.
• This hardware advancement facilitates real-time, autonomous patient monitoring, enhancing privacy and accessibility of advanced medical diagnostics.