Résumé: In recent years, medical image segmentation methods based on deep learning have been of great importance for disease diagnosis and treatment planning in clinical medicine. U-Net and a series of networks derived from it have led the research trend in medical image segmentation. In this paper, an end-to-end lightweight MLP-based medical image segmentation network G-UNeXt is proposed to address the problems of high computational complexity, large number of model parameters and slow inference in medical image segmentation networks. Firstly, this paper proposes a new skip connection method Ghost path. It reduces the semantic gap between features while combining different levels of features, and exploits the redundancy of features, which helps to improve the segmentation ability of the model for details. Secondly, this paper designs a cheap and effective G-S block, which uses low-cost linear operations to mine potential ghost features outside of intrinsic features. The G-S block reduce the number of parameters and computational complexity compared to traditional convolution. In addition, it can also adaptively calibrate the channel feature response, enhancing the characterization capability of the network and bringing some performance improvement at a lower computational cost. Finally, we build the lightweight MLP-based network G-UNeXt used Ghost path and G-S block for real-time segmentation of medical images. The results tested on the benchmark medical image segmentation datasets BUSI and ISIC2018 show that G-UNeXt reduces the parameters by 33% and the computational complexity by 23.7% compared with UNeXt. In addition, G-UNeXt also obtains faster inference speed and higher segmentation accuracy.

Langue: Anglais