.png){kind=icon}
.png){kind=icon}
Return
Multi-scale target edge uncertainty-aware consistent segmentation network
.png){kind=icon}
PRE
AI
DOI：10.1016/j.compeleceng.2024.110007.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Semi-supervised methods show promise in medical image segmentation tasks. However, most methods ignore the importance of modeling target edge regions during training, resulting the models produce many inaccurate predictions in edge regions. To solve this problem, this paper proposes a new multi-scale target edge uncertainty-aware consistency framework, called MSTNet. Specifically, a Feature Space Perturbation Branch (FSPB) module is first proposed to effectively utilize the uncertain information of the target edge to improve the semi-supervised segmentation performance. Then, a Multi-Scale Logic Gate (MSLG) module is designed to effectively capture the potential semantic and spatial information of the edge region and suppress irrelevant background noise during the encoding and decoding stages. Finally, a Multi-Scale Feature Enhancement (MSFE) module is proposed, aiming to learn more unique semantic representations to refine the segmentation of target edge regions. We compare the segmentation results of our MST-Net model and several state-of-the-art semi-supervised methods on three publicly available medical datasets. Experiments show that our method outperforms other methods in edge region segmentation. Among them, MST-Net obtained 86.16 % Dice under the 5 % setting of the LA dataset, which was 2.57 % higher than the best method.
Keywords:
Semi-supervised learning
Semi-supervised segmentation
Consistency constraint
Multi-scale extraction
Feature enhancement
Feature space perturbation
Journal
C
IF:
4.9
Papers:
6.7K
Citations:
1.3W