Semi-supervised Medical Image Classification Made Easier with Causality-Driven Learning.

Area of Science:

• Medical Image Analysis
• Machine Learning
• Causal Inference

Background:

• Semi-supervised learning (SSL) enhances medical image analysis with limited labeled data by using unlabeled data.
• Consistency regularization in SSL aligns output distributions across perturbations but can be misled by non-causal features, causing spurious results.
• Non-causal features may dominate representations, replacing intrinsic semantic information and leading to unreliable consistency.

Purpose of the Study:

• To introduce a novel causality-driven SSL framework to improve medical image classification.
• To address the issue of spurious consistency caused by non-causal features in conventional SSL methods.
• To enhance model generalization and interpretability by emphasizing causal feature extraction.

Main Methods:

• Developed a causality-driven SSL framework integrating causal inference principles.
• Employed a novel Causality-driven Consistency paradigm to refine the consistency learning process.
• Applied the framework to existing consistency regularization-based semi-supervised architectures.

Main Results:

• The proposed framework effectively reduces the impact of confounders and bias in medical image classification.
• Demonstrated superior performance in medical image classification tasks compared to conventional methods.
• Showcased enhanced generalization and interpretability of the developed model.

Conclusions:

• The causality-driven SSL framework offers a robust solution for medical image analysis, particularly with limited labeled data.
• The Causality-driven Consistency paradigm can be integrated into existing SSL architectures to boost efficiency and accuracy.
• This approach mitigates spurious consistency issues, leading to more reliable and interpretable medical image classification models.