Double-mix pseudo-label framework: enhancing semi-supervised segmentation on category-imbalanced CT volumes

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Supervised deep learning for CT segmentation requires extensive, time-consuming data labeling.
• Limited annotated data hinders the performance of current segmentation models.

Purpose of the Study:

• To improve CT volume segmentation accuracy with limited annotated data.
• To address the challenge of category-wise difficulties and data distribution in semi-supervised learning.

Main Methods:

• Proposed a confidence-difficulty weight (CDifW) allocation method to balance training across categories.
• Introduced a Double-Mix Pseudo-label Framework (DMPF) for strategic image blending based on category difficulty and distribution.
• DMPF enhances segmentation for challenging categories.

Main Results:

• Achieved a 5.1% Dice score improvement for difficult categories on 5% labeled data in the Congenital Heart Disease (CHD) dataset.
• Demonstrated a 7.0% Dice score improvement for difficult categories on 40% labeled data in the Beyond-the-Cranial-Vault (BTCV) Abdomen dataset.
• Outperformed state-of-the-art methods in semi-supervised CT segmentation.

Conclusions:

• The proposed method effectively improves segmentation performance for difficult categories in CT volumes.
• Category-wise weighting and mixture augmentation are key to enhancing semi-supervised segmentation.
• The approach is validated across datasets and significant for healthcare AI applications.