FPGA-Accelerated CNN Reconstruction for Low-Power Sparse-Array Ultrasound Imaging.

Area of Science:

• Biomedical Engineering
• Medical Imaging
• Artificial Intelligence

Background:

• Wearable ultrasound (US) imaging for real-time bladder monitoring faces challenges in power, cost, and resolution.
• High-channel-count US systems generate large data streams, complicating computational efficiency and wireless transmission.
• Existing wearable US sensors often lack the clinical accuracy and efficiency required for effective healthcare applications.

Purpose of the Study:

• To develop an algorithm-centric approach for reconstructing missing ultrasound channels using field-programmable gate array (FPGA)-accelerated deep learning.
• To reduce analog front-end requirements and power consumption in wearable ultrasound systems.
• To enable high-quality, real-time ultrasound imaging for targeted organs like the urinary bladder.

Main Methods:

• Developed a lightweight U-Net convolutional neural network (L-UNET) optimized for sparse-array radiofrequency (RF) data reconstruction.
• Deployed the L-UNET on a deep learning processing unit (DPU) using mixed quantization-aware training (Mixed-QAT) with 8-bit integer and 16-bit floating-point precision.
• Implemented a single-core FPGA accelerator for real-time inference, processing only odd-indexed physical channels and inferring even-indexed channels.

Main Results:

• Achieved a mean-squared error (MSE) of 1.48 × 10⁻², significantly lower than 32-bit floating-point precision.
• Maintained B-mode image quality with peak signal-to-noise ratio (PSNR) >18 dB and structural similarity index (SSIM) > 0.5.
• Demonstrated an average power consumption of 0.918 W in a 32-channel configuration with deterministic latency of 221 ms/frame.

Conclusions:

• Algorithm-centric reconstruction of missing US channels on embedded FPGAs is a viable strategy for wearable imaging.
• This approach effectively doubles the imaging aperture while halving analog front-end requirements and power consumption.
• The developed system offers a pathway toward fully integrated, low-power ultrasound monitoring systems for preventive healthcare and early disease diagnosis.