Image-based deep causal discovery for explaining medical images with surgical knowledge

Area of Science:

• Medical Imaging
• Machine Learning
• Causal Inference

Background:

• Machine learning models in medicine often lack interpretability, hindering comprehension of decision-making processes.
• Current saliency maps visualize image regions but do not capture causal relationships or their direction.
• Existing causal discovery methods lack human-interpretable visualization for image data.

Purpose of the Study:

• To develop a deep causal discovery model for visualizing inherent causal relationships within medical images.
• To improve the interpretability of machine learning models in medical applications, particularly in surgical planning.
• To address the challenge of extracting and visualizing causal relationships from image data in a human-understandable format.

Main Methods:

• Proposed a deep causal discovery model incorporating L1 norm regularization on the causal matrix.
• Integrated spatial information from image patches into the model.
• Visualized causal relationships as interpretable causal graphs between image patches.

Main Results:

• The model successfully generated sparse and interpretable causal graphs for medical images.
• Demonstrated visualization of causal relationships between mandibular features and surgical planning.
• Applied to a mandibular reconstruction planning database, showing explainability of surgical knowledge in images.

Conclusions:

• The proposed deep causal discovery model effectively visualizes causal relationships in medical images.
• This approach enhances the interpretability of machine learning in surgical planning.
• Facilitates systematization of diagnosis and treatment processes by visualizing surgeon decision-making.