Accelerated inference for thyroid nodule recognition in ultrasound imaging using FPGA.

Area of Science:

• Medical Imaging
• Computer Engineering
• Endocrinology

Background:

• Thyroid cancer incidence is rising, necessitating efficient diagnostic tools.
• Current Central Processing Units (CPUs) and Graphics Processing Units (GPUs) have limitations in speed, energy consumption, and cost for medical image analysis.
• Existing hardware is inadequate for future green, efficient, and accessible healthcare diagnostics.

Purpose of the Study:

• To propose an efficient quantized inference method using a Field-Programmable Gate Array (FPGA) for thyroid nodule identification.
• To overcome the limitations of traditional CPUs and GPUs in processing speed, energy consumption, and scalability.
• To enhance the speed, accuracy, and environmental sustainability of ultrasound image analysis.

Main Methods:

• Utilized the YOLOv4-tiny neural network model on a ZYNQ7020 FPGA platform.
• Implemented K-means++ optimization for software performance enhancement.
• Applied 8-bit weight quantization, batch normalization, and convolutional layer fusion for hardware performance improvement.

Main Results:

• Achieved an average accuracy of 81.44% on the Tn3k dataset and 81.20% on an internal test set.
• FPGA power consumption was 3.119 W, significantly lower than CPU (45 W) and GPU (68 W).
• FPGA demonstrated 17.6 times higher energy efficiency than CPU and comparable efficiency to high-end GPUs, with significantly lower power usage.

Conclusions:

• FPGA-based quantized inference offers a novel, efficient solution for medical imaging diagnostics, particularly for thyroid nodule identification.
• This approach significantly outperforms CPUs in speed and GPUs in power consumption and energy efficiency.
• Mid-to-low-end FPGAs can achieve performance comparable to commercial GPUs, paving the way for sustainable and accessible medical imaging analysis.