Self-Supervised Learning for Few-Shot Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Fully-supervised deep learning segmentation models struggle with new classes and require extensive annotated data for fine-tuning.
• Few-shot semantic segmentation (FSS) addresses this by learning from few examples without fine-tuning, but current methods are ill-suited for annotation-scarce medical imaging.
• Existing FSS methods for natural images rely on large annotated datasets, a limitation in medical contexts.

Purpose of the Study:

• To develop a novel self-supervised few-shot semantic segmentation (FSS) framework for medical images, bypassing the need for training annotations.
• To improve the adaptability and efficiency of FSS in resource-constrained medical imaging scenarios.
• To enhance segmentation accuracy for unseen classes in medical images.

Main Methods:

• Proposed a self-supervised FSS framework (SSL-ALPNet) for medical images, eliminating the need for manual annotations during training.
• Utilized superpixel-based pseudo-labels to generate supervision signals for the self-supervised learning process.
• Introduced an adaptive local prototype pooling module to enhance prototype networks and boost segmentation performance.

Main Results:

• The proposed SSL-ALPNet framework demonstrated general applicability across diverse medical imaging tasks, including abdominal organ segmentation (CT/MRI) and cardiac MRI segmentation.
• Achieved higher Dice scores compared to conventional FSS methods that necessitate manual annotations for training.
• Successfully bypassed the requirement for extensive annotations, proving effective in annotation-scarce medical imaging scenarios.

Conclusions:

• The novel self-supervised FSS framework (SSL-ALPNet) effectively addresses the limitations of traditional supervised methods in medical image segmentation.
• The approach offers a practical solution for segmentation tasks where annotated data is scarce, demonstrating strong performance across multiple medical imaging modalities and tasks.
• SSL-ALPNet provides a significant advancement in few-shot semantic segmentation for medical applications, outperforming existing methods that rely on manual annotations.