One-shot neuroanatomy segmentation through online data augmentation and confidence aware pseudo label

Area of Science:

• Medical Image Analysis
• Deep Learning
• Neuroimaging

Background:

• Supervised deep learning methods dominate brain segmentation but require extensive labeled data.
• One-shot segmentation, learning from minimal labeled data, presents a significant challenge.

Purpose of the Study:

• To develop an efficient end-to-end deep learning network for one-shot brain segmentation.
• To address the limitations of data-intensive supervised approaches by leveraging unlabeled data.

Main Methods:

• A unified network integrating deformation modeling and segmentation with a shared encoder.
• Multi-scale feature extraction and coarse-to-fine deformation field estimation.
• Online data augmentation and confidence-aware pseudo-labeling for unlabeled data.

Main Results:

• The proposed network significantly outperforms deep single-atlas and traditional multi-atlas segmentation methods.
• Robust segmentation performance demonstrated across three benchmark datasets, including multi-center data.
• Effective handling of appearance variations through the shared encoder.

Conclusions:

• The developed one-shot segmentation network offers a powerful and data-efficient alternative for brain image analysis.
• The approach shows promise for clinical applications where labeled data is scarce.
• The method's robustness across diverse datasets highlights its potential for real-world use.