Generating Counterfactual Explanations For Causal Inference in Breast Cancer Treatment Response.

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Oncology
• Causal Inference in Healthcare

Background:

• Radiomics from MRI shows promise in predicting breast cancer treatment response to neoadjuvant systemic therapy (NST).
• Current machine learning models lack expert-level interpretability and causal inference capabilities.
• Understanding causal relationships is crucial for guiding treatment strategies and gaining clinical acceptance.

Purpose of the Study:

• To extract causal relationships between imaging phenotypes, clinical information, molecular features, and treatment response after NST.
• To enhance the interpretability of machine learning models in predicting breast cancer treatment outcomes.
• To leverage counterfactual explanations for causal inference in oncology.

Main Methods:

• Utilized counterfactual explanations to identify causal links.
• Applied methodology to a publicly available breast cancer dataset.
• Compared counterfactual explanations with traditional methods like LIME and Shapley.

Main Results:

• Demonstrated the extraction of causal relationships using counterfactual explanations.
• Identified key imaging phenotypes, clinical information, and molecular features influencing treatment response.
• Showcased the potential of counterfactual explanations for deeper insights into treatment prediction.

Conclusions:

• Counterfactual explanations offer a powerful approach for causal inference in breast cancer treatment response prediction.
• This methodology enhances the interpretability of radiomics-based machine learning models.
• Findings can inform personalized treatment strategies and clinical decision-making in oncology.