Attentional adversarial training for few-shot medical image segmentation without annotations

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Medical image segmentation is crucial for clinical research but suffers from a lack of annotated data.
• Deep neural networks require extensive labeled data, hindering robust model training for segmentation tasks.
• Few-shot learning offers a solution by enabling predictions for new classes with minimal annotations.

Purpose of the Study:

• To propose a novel few-shot semantic segmentation framework, the prototype-based generative adversarial network (PG-Net), for medical image segmentation without requiring extensive annotations.
• To enhance the quality and generalization ability of medical image segmentation models in low-data regimes.

Main Methods:

• Developed a prototype-based generative adversarial network (PG-Net) comprising a prototype-based segmentation network (P-Net) and a guided evaluation network (G-Net).
• P-Net extracts multi-scale features and local spatial information for refined predictions.
• G-Net utilizes an attention mechanism to refine segmentation mask distributions, guided by an adversarial training strategy.

Main Results:

• PG-Net demonstrated enhanced segmentation quality through adversarial training.
• Comparative experiments showed PG-Net offers superior robustness and generalization compared to state-of-the-art few-shot segmentation methods.
• The framework performed effectively on both abdominal Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) datasets.

Conclusions:

• The proposed PG-Net effectively addresses the challenge of limited annotated data in medical image segmentation.
• PG-Net achieves robust and prominent generalization capabilities across diverse medical imaging modalities.
• This framework holds significant potential for advancing clinical research through improved automated image analysis.