Beyond Known Reality: Exploiting Counterfactual Explanations for Medical Research

TL;DR

This work addresses the need for human-friendly explanations in AI-assisted medical decision-making by applying counterfactual explanations to MRI-derived features for pediatric posterior fossa tumors. It develops a DiCE-based framework to generate plausible, minimal-feature changes that would alter tumor-type predictions, enabling both classification guidance and interpretation of feature importance. The study demonstrates that counterfactuals can support data augmentation, reveal key MRI features driving distinctions, and integrate into a distance-based decision space to aid clinicians. Although limited by dataset size, the approach shows promise for personalized, explainable diagnostic insights and motivates future scaling to broader MRI protocols and diseases. Overall, counterfactual explanations offer a human-centric lens to understand and validate AI-driven predictions in pediatric neuro-oncology, potentially reducing invasive procedures and informing treatment planning.

Abstract

The field of explainability in artificial intelligence (AI) has witnessed a growing number of studies and increasing scholarly interest. However, the lack of human-friendly and individual interpretations in explaining the outcomes of machine learning algorithms has significantly hindered the acceptance of these methods by clinicians in their research and clinical practice. To address this issue, our study uses counterfactual explanations to explore the applicability of "what if?" scenarios in medical research. Our aim is to expand our understanding of magnetic resonance imaging (MRI) features used for diagnosing pediatric posterior fossa brain tumors beyond existing boundaries. In our case study, the proposed concept provides a novel way to examine alternative decision-making scenarios that offer personalized and context-specific insights, enabling the validation of predictions and clarification of variations under diverse circumstances. Additionally, we explore the potential use of counterfactuals for data augmentation and evaluate their feasibility as an alternative approach in our medical research case. The results demonstrate the promising potential of using counterfactual explanations to improve AI-driven methods in clinical research.

Figures (5)

• Figure 1: The figure illustrates a hypothetical clinical scenario showcasing the practical application of counterfactuals. The data consists of features extracted from provided multi-parametric brain MRIs, rather than the raw images themselves.
• Figure 2: Generating counterfactual explanations on tumor types. The left illustrates feature manipulation using an ML model with a counterfactual approach. The right delves deeper into the process and counterfactual explanation concept, exemplified by two tumor types.
• Figure 3: An overview of our approach demonstrating the generation of counterfactual explanations.
• Figure 4: The original data distributions for MB-PA and BG-PA, highlighting key features for MB-PA and their behavior when applied to BG-PA. The figure demonstrates how the top 5 features vary significantly between MB and PA, while BG and PA distributions remain almost identical, lacking discernible features.
• Figure 5: The distributions of the original data and the generated data.