Beyond Correlation: Causal Intervention for Multi-Label Medical Image Diagnosis

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Causal Inference

Background:

• Multi-disease diagnosis is crucial in clinical practice, yet current deep learning models often focus on single diseases.
• Existing multi-label methods rely on correlations, failing to capture true causal relationships and leading to diagnostic inaccuracies.
• Spurious feature-disease associations arise from co-occurring conditions, hindering AI diagnostic performance and interpretability.

Purpose of the Study:

• To develop a novel framework integrating causal reasoning into multi-label medical image diagnosis.
• To enable AI models to identify true causal signals, overcoming limitations of correlation-based inference.
• To enhance the accuracy and interpretability of AI-assisted diagnosis for concurrent diseases.

Main Methods:

• Proposed a framework incorporating causal intervention for multi-label medical image diagnosis.
• Modeled latent disease-related confounders and applied backdoor adjustment to disentangle causal effects.
• Learned shared feature representations as confounding variables to refine image-derived features.

Main Results:

• The causal framework consistently outperformed existing approaches across four diverse medical imaging datasets (ODIR, LID-FFA, Endo, Chestpert).
• Demonstrated effective separation of diagnoses for co-occurring diseases.
• Showcased improved accuracy and interpretability in multi-disease diagnostic tasks.

Conclusions:

• Causal reasoning significantly enhances the reliability and clinical applicability of AI-assisted diagnosis.
• The proposed method offers a robust solution for the complex challenge of multi-disease diagnosis.
• Future AI diagnostic systems can benefit from incorporating causal inference to address confounding factors.