SCGAN: Sparse CounterGAN for Counterfactual Explanations in Breast Cancer Prediction.

Area of Science:

• Radiomics and Medical Imaging
• Computational Biology and Bioinformatics
• Machine Learning and Artificial Intelligence

Background:

• Radiomics from MRI show promise in predicting breast cancer treatment response to neoadjuvant systemic therapy (NST).
• Understanding causal links between Imaging, Clinical, and Molecular (ICM) features and treatment response is crucial for personalized medicine.
• Existing counterfactual explanation methods face challenges with high dimensionality, realism, and feature confounding.

Purpose of the Study:

• To propose a novel method, Sparse CounteRGAN (SCGAN), for generating realistic and interpretable counterfactual explanations.
• To reveal causal relationships between ICM features and treatment response in breast cancer patients undergoing NST.
• To address limitations of existing counterfactual generation techniques.

Main Methods:

• Developed SCGAN, a generative approach learning data distribution for realistic counterfactual instance generation.
• Incorporated dropout training for the discriminator to enforce sparsity in counterfactuals.
• Introduced a diversity term in the loss function to maximize distances between generated counterfactuals.

Main Results:

• SCGAN successfully generates sparse and diverse counterfactual instances.
• The generated counterfactuals demonstrate plausibility and feasibility.
• Evaluated on multiple datasets, SCGAN outperforms existing methods in revealing causal relationships.

Conclusions:

• SCGAN provides a valuable tool for causal inference in high-dimensional medical data.
• The method enhances understanding of ICM features' impact on breast cancer treatment response.
• SCGAN facilitates more informed clinical decision-making for breast cancer management.