.png){kind=icon}
.png){kind=icon}
Return
Adaptive Noise Augmentation and Dual-Branch Learning for Semi-Supervised Medical Image Segmentation
.png){kind=icon}
PRE
AI
K
R
J
Y
Y
Y
DOI：10.1002/ima.70398.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In the field of medical image segmentation, semi-supervised learning is widely recognized as a method with tremendous development potential. However, common consistency regularization methods can introduce fixed noise, leading to pixel point confusion of different categories within high-density pixel regions in images, thereby affecting the accuracy of target edge segmentation. Furthermore, accurately identifying target boundary regions during labeled data training and recognizing high-confidence areas during unlabeled data training are crucial for the model's learning. In this paper, we introduce an Adaptive Noise Method that automatically identifies and adds adaptive noise to different image regions, improving pixel distribution accuracy. We also propose a Dual-Branch Edge Reinforcement Learning method, which incorporates multiple learning paradigms to enhance edge inference and correct perceptual biases. Furthermore, we present a Dual-Branch Co-Decision Strategy for precise localization of high-confidence regions in unlabeled data, integrating multiple prediction branches for improved accuracy. Our proposed AR-Net, a consistency network, outperforms existing methods on three medical image datasets, demonstrating superior segmentation performance. Notably, on the challenging Pancreas-CT dataset, AR-Net achieves a Dice score of 80.79% and a Jaccard index of 68.28%, showing clear performance gains under the same semi-supervised setting. The source code will be made publicly available at https://github.com/IRSI-SK/AR-Net.
Keywords:
2D/3D CT/MRI
adaptive noise augmentation
dual-branch co-decision strategy
dual-branch learning
medical image segmentation
semi-supervised learning
Journal
IF:
2.5
Papers:
2.1K
Citations:
2.3K
Organization
Cited Papers