LKSNeXt: An Efficient Medical Image Segmentation Network with Large Kernels and Lightweight Structure

Abstract

U-shaped architectures play a critical role in medical image segmentation. Traditional fully convolutional U-shaped networks, however, encounter numerous challenges in processing medical images, particularly in capturing long-range dependencies and global contextual information.Recently, hybrid architectures combining convolutional neural networks (CNNs) and Transformers have shown potential advantages in addressing these issues, though their high computational complexity and resource demands limit their widespread application.In contrast, lightweight networks demonstrate significant advantages in processing medical data, particularly in resource-limited settings. To address these challenges, we introduce a new network architecture named LKSNeXt, which effectively extracts global information through inductive biases, significantly enhancing the efficiency and accuracy of medical image segmentation.LKSNeXt employs large kernels and an inverted bottleneck design to thoroughly mix long-range spatial and positional information, effectively extracting global context. The network enhances its capability to process remote dependencies through a specifically designed structure, thereby more efficiently capturing and utilizing global context.Additionally, an innovatively introduced skip-connection fusion module further enhances feature integration, improving the model’s robustness and expressiveness. Experimental results across multiple datasets demonstrate that LKSNeXt significantly improves segmentation precision and efficiency under constraints of increased inference speed and reduced model size, showing strong potential for practical application.