Efficient few-shot medical image segmentation via self-supervised variational autoencoder.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Few-shot medical image segmentation commonly employs joint registration and segmentation models.
• Spatial misalignments in registration can compromise segmentation accuracy and quality.

Purpose of the Study:

• To develop an end-to-end model, EFS-MedSeg, for enhanced few-shot medical image segmentation.
• To address inaccuracies caused by registration misalignments using data augmentation and self-supervised learning.

Main Methods:

• EFS-MedSeg utilizes two labeled atlases and few unlabeled images.
• Employs a 3D random regional switch strategy for atlas augmentation to improve generalization and prevent overfitting.
• Incorporates a variational autoencoder for weighted reconstruction and a self-contrastive module for feature extraction guided by anatomical priors.

Main Results:

• EFS-MedSeg achieves performance comparable to fully-supervised methods on multi-modal medical image datasets.
• Demonstrates superior Dice scores over the second-best method by 1.4% (OASIS), 9.1% (BCV), and 1.1% (BCH).
• Highlights robustness and adaptability across diverse medical imaging datasets.

Conclusions:

• EFS-MedSeg offers a robust and accurate solution for few-shot medical image segmentation.
• The model's enhancements in data augmentation and self-supervised learning contribute to improved segmentation accuracy and boundary smoothness.
• The approach shows significant potential for advancing medical image analysis in low-data regimes.