A Lightweight Medical Image Segmentation Network Based on Fusion of Local and Global Features

Due to the variability in the positions, sizes, and shapes of regions in medical image segmentation, most existing automatic segmentation tasks remain highly challenging. Traditional convolutional neural networks (CNNs) have inherent limitations in capturing global contextual information due to their fixed receptive field sizes. While Transformer-based architectures with self-attention mechanisms can effectively model long-range dependencies, they come with high computational complexity, making them less practical for some applications. To address these challenges, we propose a novel dual-branch network called the Global and Local Feature Fusion Network (GLFFNet). This architecture leverages the strengths of both CNNs and Transformers by fusing global and local features to capture both fine-grained details and broad contextual information. Additionally, we employ a cross-fusion approach within the network’s decoder to further enhance feature integration and segmentation accuracy. Comprehensive comparative experiments are conducted on two benchmark segmentation datasets, ISIC2018 and Kvasir. The experimental results convincingly demonstrate that our proposed model achieves state-of-the-art performance, striking an optimal balance between parameter count, computational complexity, and segmentation accuracy. These findings underscore the potential of GLFFNet in advancing the field of medical image segmentation.