Comparative Analysis of Pure Convolution Narrow U-Net for Colorectal Polyp Segmentation on GPU, CPU and FPGA

Addressing the task of manually segmenting unhealthy portions from healthy areas in biomedical images, particularly for colorectal polyps, poses significant challenges in terms of time and labor. There is a need for expedited and reliable diagnosis to facilitate prompt and appropriate patient treatment. This research focuses on optimizing deep learning algorithms for the segmentation of colorectal polyp images, specifically targeting low-power platforms with minimal resource utilization. In this paper we have introduced a Pure Convolution Narrow U-Net (PCN U-Net) for polyp image segmentation. The proposed PCN U-Net achieves a mean Intersection over Union (mIoU) score of 0.90 while demonstrating a throughput of 32 frames per second (FPS) when evaluated with CVC-ClinicDB dataset on a GPU platform. Our proposed architecture is implemented and assessed across three distinct platforms: CPU, GPU, and FPGA. The results demonstrate superior performance on FPGA, showcasing a remarkable energy efficiency of 10.778 FPS/Watt and 10.585 FPS/Watt for CVC-ClinicDB and Kvasir-Seg datasets, respectively with the lowest power consumption recorded at 3.2 Watt.