Dual contextual learning for semi-supervised medical image classification

Area of Science:

• Artificial Intelligence
• Medical Image Analysis
• Computer Vision

Background:

• Semi-supervised learning (SSL) is crucial for medical image classification due to limited labeled data.
• Existing pseudo-labeling methods struggle with ambiguous cases, leading to error accumulation.
• Medical images possess rich contextual information that can improve classification accuracy.

Purpose of the Study:

• To propose a Hierarchical Semantic Calibration (HSC) framework to enhance pseudo-labeling reliability in medical image classification.
• To leverage contextual relationships within medical image data for more robust supervision.
• To improve accuracy in classifying medical images, especially in challenging cases with limited annotations.

Main Methods:

• Introduced a local semantic neighborhood alignment module to enforce consistency among k-nearest neighbors.
• Implemented a global cluster-prototype calibration module using contrastive learning for class-level representation alignment.
• Developed a neighborhood-prototype consistency regularization to bridge local and global scales adaptively.

Main Results:

• Achieved 92.24% accuracy on NCT-CRC-HE with only 200 labeled samples, outperforming existing methods.
• Attained 94.17% accuracy on ISIC2018 using 20% labeled data, demonstrating significant improvement.
• HSC consistently outperformed state-of-the-art methods in medical image classification tasks.

Conclusions:

• The proposed HSC framework effectively enhances pseudo-labeling reliability by utilizing hierarchical semantic information.
• HSC demonstrates superior performance in medical image classification, particularly with limited labeled data.
• The method offers a robust solution for accurate disease pattern recognition despite imaging variations and ambiguous features.