Robust Prototypical Few-Shot Organ Segmentation With Regularized Neural-ODEs.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Deep learning for image semantic segmentation demands extensive annotated data, which is costly in medical domains.
• Few-Shot Learning (FSL) addresses this by enabling generalization to new classes with minimal annotations.

Purpose of the Study:

• To introduce Regularized Prototypical Neural Ordinary Differential Equation (R-PNODE) for Few-Shot Segmentation (FSS) in medical imaging.
• To enhance the performance and adversarial robustness of FSS methods.

Main Methods:

• Leveraging Neural Ordinary Differential Equations (Neural-ODEs) with cluster and consistency losses.
• Constraining support and query features within the same classes in the representation space.
• Evaluating R-PNODE on multi-organ segmentation datasets in in-domain and cross-domain FSS settings.

Main Results:

• R-PNODE significantly outperforms existing Convolutional Neural Network (CNN) based FSS methods.
• The method demonstrates increased adversarial robustness against various common attacks.
• Superior performance is observed across different attack intensities and designs.

Conclusions:

• R-PNODE offers an effective solution for medical Few-Shot Segmentation with reduced annotation requirements.
• The proposed method provides enhanced robustness against adversarial attacks compared to traditional CNN-based approaches.
• R-PNODE shows significant potential for improving medical image analysis workflows.