Accelerated inference for thyroid nodule recognition in ultrasound imaging using FPGA

Thyroid cancer is the most prevalent malignant tumour in the endocrine system, with its incidence steadily rising in recent years. Current central processing units (CPUs) and graphics processing units (GPUs) face significant challenges in terms of processing speed, energy consumption, cost, and scalability in the identification of thyroid nodules, making them inadequate for the demands of future green, efficient, and accessible healthcare. To overcome these limitations, this study proposes an efficient quantized inference method using a field-programmable gate array (FPGA). We employ the YOLOv4-tiny neural network model, enhancing software performance with the K-means + + optimization algorithm and improving hardware performance through techniques such as 8-bit weight quantization, batch normalization, and convolutional layer fusion. The study is based on the ZYNQ7020 FPGA platform. Experimental results demonstrate an average accuracy of 81.44% on the Tn3k dataset and 81.20% on the internal test set from a Chinese tertiary hospital. The power consumption of the FPGA platform, CPU (Intel Core i5-10200 H), and GPU (NVIDIA RTX 4090) were 3.119 watts, 45 watts, and 68 watts, respectively, with energy efficiency ratios of 5.45, 0.31, and 5.56. This indicates that the FPGA's energy efficiency is 17.6 times that of the CPU and 0.98 times that of the GPU. These results show that the FPGA not only significantly outperforms the CPU in speed but also consumes far less power than the GPU. Moreover, using mid-to-low-end FPGAs yields performance comparable to that of commercial-grade GPUs. This technology presents a novel solution for medical imaging diagnostics, with the potential to significantly enhance the speed, accuracy, and environmental sustainability of ultrasound image analysis, thereby supporting the future development of medical care.