.png){kind=icon}
.png){kind=icon}
Return
Lite-MixedNet: Lightweight and efficient hybrid network for medical image segmentation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.patcog.2025.111378.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In recent years, the field of medical image segmentation has seen gains in performance by combining Convolutional Neural Network (CNN), which capture local information, with Vision Transformer (ViT), which capture global information. However, while ViT brings about performance improvements, it also increases the computational complexity of the network. Furthermore, this paper identifies that, due to the limited size of medical image datasets, the self-attention mechanism's approach to capturing global information through Query (Q), Key (K), and Value (V) is not the optimal choice. Therefore, this paper proposes a lightweight and effective segmentation network called Lite-MixedNet. Specifically, Lite-MixedNet reduces the parameters and introduces two new self-attention mechanisms: Efficient-and-simple Attention (Es Attention) and Channel Efficient-and-simple Attention (C-Es Attention). Es Attention achieves global modeling through Q and V while enhancing the differentiation of spatial information, thereby reducing the computational complexity of the network and improving segmentation accuracy. C-Es Attention, building on Es Attention, proposes the Channel Feature Selection (CFS) mechanism, which performs feature learning at the channel dimension and eliminates irrelevant or weakly relevant information, thereby focusing on key information. The paper conducts experiments on three different types of datasets to validate the feasibility and effectiveness of our method. Our code will be available at https://github.com/vpsg-research/Lite-MixedNet.
Keywords:
Medical image segmentation
Convolutional Neural Network
Vision Transformer
Self-attention mechanism
Journal
IF:
7.6
Papers:
1.3W
Citations:
4.5W
Organization
No organization information available
Cited Papers