Real-Time Endoscopic Image Semantic Segmentation for Efficient Polyp Identification

The advancement of endoscopic procedures has significantly enhanced diagnostic capabilities in polyp identification and removal, which is critical for colorectal cancer prevention. However, the success of these procedures hinges on the accuracy and speed of real-time image analysis. Traditional image processing methods often fail to provide timely and precise segmentation of endoscopic images. This paper presents the implementation of endoscopic image semantic segmentation using U-Net-like convolutional neural networks (CNNs) and depicts how it can be implemented efficiently on field-programmable gate arrays (FPGAs). The proposed methodology involves a structured pipeline for model development, data preparation, and deployment on FPGA hardware to ensure optimal resource utilization and real-time performance. Experimental results demonstrate that our CNN model achieves a Dice coefficient of 0.92 and an Intersection over Union (IoU) score of 0.85. Furthermore, the real-time FPGA implementation achieved an average inference time of 0.007 seconds per image and a speedup of approximately 2.5× compared to NVIDIA RTX 3070 Ti GPU implementation. Additionally, the FPGA solution consumes approximately 31× less power than the GPU.