Optimized Hybrid Deep Learning‐Based FPGA Accelerators for Denoising of Ultrasound Breast Images

This paper introduces a novel image‐denoising technique that integrates a hybrid deep learning (DL) model with a self‐improved orca predation (SOP) strategy. This hybrid model improves denoising performance by integrating a Convolutional Neural Network (CNN) with Bidirectional Long Short‐Term Memory (Bi‐LSTM). The hybrid model’s hyperparameters are enhanced using the SOP technique, resulting in superior denoising outcomes. The proposed approach is experimentally tested with the INbreast and CBIS‐DDSM datasets. The results demonstrate that the suggested method outperforms conventional approaches, making it a viable option for image‐denoising applications. The suggested technique achieved a PSNR of 35.905 on the INbreast dataset and 37.08 on the CBIS‐DDSM dataset. However, the DL model demands a significant amount of memory and computing capacity, limiting its implementation on edge devices and causing computing delays and energy loss. Field Programmable Gate Arrays (FPGAs) are ideal for practical applications due to their high computational capability and low power consumption. In this paper, we implement the proposed model on a ZCU104 FPGA board, evaluate its performance, and analyze resource utilization. The experimental outcome shows that the proposed network on the chosen FPGA achieves an impressive execution time of 4.25 s, low power consumption of 3.2 W, and a throughput of 47 images per second.