UC-Hybrid: Uncertainty-based contrastive learning on hybrid network for medical image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning significantly advanced medical image segmentation, but struggles with small object detection and inter-class accuracy bias in clinical applications.
• Existing methods often face a trade-off between segmenting small objects and minimizing false positives.

Purpose of the Study:

• To propose an uncertainty-based contrastive learning technique (UncerNCE) and a hybrid deep learning architecture.
• To enhance the segmentation performance for small organs in multi-organ segmentation tasks.
• To address inter-class accuracy bias and reduce noise (false positives).

Main Methods:

• Developed a hybrid backbone network combining convolutional and transformer layers.
• Implemented an uncertainty-based contrastive learning approach (UncerNCE) for focused learning on uncertain regions.
• Utilized spotlight learning based on uncertainty to improve segmentation of all classes.

Main Results:

• Achieved superior segmentation performance for small organs compared to state-of-the-art methods.
• Successfully addressed the multi-class accuracy bias, improving performance across all organ classes.
• Demonstrated effective noise suppression, reducing false positives while maintaining high segmentation accuracy.
• Validated results on BTCV and 1K datasets.

Conclusions:

• The proposed UncerNCE method and hybrid architecture significantly improve medical image segmentation, particularly for small organs.
• Uncertainty-based contrastive learning effectively resolves trade-offs between small object segmentation and noise reduction.
• This approach offers a promising solution for accurate and robust multi-organ segmentation in clinical settings.