Towards Causal Explainable AI in Cancer Diagnosis: Advances, Challenges, and Future Directions

Area of Science:

• Oncology
• Artificial Intelligence (AI)
• Causal Inference
• Explainable AI (XAI)

Background:

• AI is revolutionizing oncology with improved cancer detection, diagnosis, and personalized treatments.
• Clinical adoption of AI is limited by the opacity of 'black-box' models, raising concerns about reliability, accountability, and ethics.
• Explainable AI (XAI) offers solutions but faces challenges like low fidelity, fairwashing, and susceptibility to bias and spurious correlations.

Purpose of the Study:

• To provide the first comprehensive review of Causal Explainable AI (CXAI) specifically for cancer diagnosis.
• To address the shortcomings of current AI and XAI methods in oncology by integrating causal inference.
• To explore current applications, identify challenges, and propose future directions for CXAI in cancer diagnosis.

Main Methods:

• Comprehensive literature review focusing on CXAI applications in oncology.
• Synthesis of existing studies on causality, explainability, and AI in healthcare, with a focus on cancer diagnosis.
• Analysis of CXAI's potential to improve model robustness, fairness, and clinical relevance.

Main Results:

• CXAI leverages causal inference to overcome limitations of traditional XAI, enhancing AI model robustness, fairness, and clinical relevance in oncology.
• The review summarizes CXAI applications across various cancer types, highlighting its potential to make AI-driven oncology more transparent and effective.
• CXAI improves clinical decision-making by capturing causal relationships in medical data, enhancing predictive accuracy and aligning AI insights with clinical reasoning.

Conclusions:

• CXAI represents a significant advancement, offering a more trustworthy and transparent approach to AI in cancer diagnosis.
• By increasing transparency and trust, CXAI facilitates the adoption of AI in oncology, leading to informed diagnostic decisions and personalized patient care.
• Future research should focus on further developing and validating CXAI methodologies to fully realize their potential in clinical oncology.