Semi-Supervised Medical Image Segmentation Based on Deep Consistent Collaborative Learning.

Area of Science:

• Medical Image Analysis
• Artificial Intelligence
• Machine Learning

Background:

• High cost of labeled data limits medical image analysis.
• Semi-supervised learning leverages unlabeled data to overcome label scarcity.
• Existing methods may not fully exploit diverse deep learning architectures.

Purpose of the Study:

• To develop a novel semi-supervised medical segmentation framework, DCCLNet.
• To address the challenge of limited labeled data in medical imaging.
• To integrate consistency learning and collaborative training for improved segmentation accuracy.

Main Methods:

• DCCLNet employs deep consistent co-learning with feature and input perturbations.
• Feature perturbation uses auxiliary decoders to enhance CNN backbone robustness.
• Input perturbation utilizes a mean teacher architecture for guided learning.
• Collaborative training integrates Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs).

Main Results:

• Achieved Dice coefficients of 0.890 on the ACDC dataset and 0.812 on the Prostate dataset.
• Demonstrated the effectiveness of individual components through ablation studies.
• Showcased improved segmentation accuracy through synergistic CNN-ViT collaboration.

Conclusions:

• DCCLNet offers a robust semi-supervised approach for medical image segmentation.
• The framework effectively mitigates the impact of limited labeled data.
• Integrating consistency learning and collaborative training enhances segmentation performance.