Semi-Supervised Medical Image Segmentation Using Adversarial Consistency Learning and Dynamic Convolution Network.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Machine Learning

Background:

• Semi-supervised medical image segmentation networks often struggle with inaccurate supervision from unlabeled data.
• Existing methods typically rely on single-level consistency learning, ignoring data relationships and leading to uncertain predictions.
• These networks often have large parameter counts and high overfitting risks due to small training datasets.

Purpose of the Study:

• To propose a novel adversarial self-ensembling network (ASE-Net) for improved semi-supervised medical image segmentation.
• To enhance prediction quality by effectively utilizing both labeled and unlabeled data.
• To reduce computational costs and memory overhead compared to existing methods.

Main Methods:

• Introduced an adversarial consistency training strategy (ACTS) using two discriminators for pixel-level and image-level consistency learning.
• Developed a dynamic convolution-based bidirectional attention component (DyBAC) for adaptive weight adjustment based on input sample structure.
• Integrated ACTS and DyBAC into the proposed ASE-Net architecture.

Main Results:

• ASE-Net demonstrated superior performance compared to state-of-the-art networks on three public datasets.
• The proposed methods effectively improved prediction quality by capturing relationships between labeled and unlabeled data.
• Reduced computational costs and memory overhead were observed.

Conclusions:

• ASE-Net offers a promising solution for semi-supervised medical image segmentation.
• The combination of ACTS and DyBAC enhances feature representation and mitigates overfitting.
• The approach provides a more efficient and accurate method for medical image segmentation tasks.