Explainable and Trustworthy AI in Oncology: A Systematic Review Across Major Organ Systems

Résumé fourni par la source

The increasing adoption of AI in cancer diagnosis requires transparency and interpretability, particularly in AI-based clinical decision support systems. However, the inherent complexity and opacity of many AI models hinder their clinical acceptance and integration into healthcare workflows. This systematic review provides a comprehensive overview of eXplainable AI (XAI) techniques applied in the diagnosis of cancer in nine major organs, including the lungs, skin, brain, breast, gastrointestinal, colorectal, kidney, prostate, and ovary. The study systematically reviews and compares widely used XAI methods, covering both model-specific and model-agnostic approaches, and analyzes them based on interpretability aspects such as global vs. local explanations and intrinsic vs. post-hoc transparency, and provides guidance for selecting appropriate XAI techniques for specific diagnostic applications. Recent studies employing XAI in medical imaging and predictive modeling are analyzed to identify trends in modality usage, evaluation metrics, and methodological choices. Comparative insights into how different XAI techniques perform across various organ systems are presented in structured tables. This critical comparison highlights the strengths and limitations of these methods in real-world diagnostic contexts, noting challenges such as clinical trust, evaluation inconsistency, and integration into existing workflows. The review critically discusses the gap between advancements and their practical clinical utility, underscoring the urgent need for clinically relevant validation. In addition, current gaps are identified, and future research directions are discussed to improve the utility and reliability of XAI. The study provides a foundational reference for researchers and clinicians, supporting informed decisions and facilitating the effective adoption of XAI in oncology practice. This systematic review was conducted using a structured search strategy, predefined eligibility criteria, and PRISMA-guided study selection to ensure a transparent and reproducible synthesis of the literature.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.