CausalMixNet: A mixed-attention framework for causal intervention in robust medical image diagnosis

Area of Science:

• Artificial Intelligence
• Medical Imaging Analysis
• Causal Inference

Background:

• Confounding factors in medical images hinder deep learning accuracy and generalization.
• Unobservable confounders pose a significant challenge for causal exploration in AI models.

Purpose of the Study:

• To introduce CausalMixNet, a novel methodology for probing causal relationships in medical images.
• To mitigate unobservable confounding factors and enhance mediator learning for causal intervention.

Main Methods:

• CausalMixNet utilizes query-mixed intra-attention and key&value-mixed inter-attention.
• Integrates non-local reasoning module (NLRM) and key&value-mixed inter-attention (KVMIA) for front-door adjustment.
• Employs patch-masked ranking module (PMRM) and query-mixed intra-attention (QMIA) for mediator learning.

Main Results:

• CausalMixNet achieves superior accuracy and F1-scores compared to existing methods.
• Demonstrates an average improvement of 3% over the closest competitor across datasets.
• Shows robust performance against noise, gender bias, and attribute bias.

Conclusions:

• CausalMixNet effectively handles unobservable confounders in medical image analysis.
• The proposed method maintains stable performance in challenging conditions, improving causal inference.
• Patch mixing mechanism enhances lesion-related features and average causal effect inference.