Explainable Deep Learning for Imaging‐Based Skin Lesion Diagnosis: A Systematic Literature Review

Abstract

In the latest years, the use of Deep Learning (DL) in imaging‐based skin lesion diagnosis has become increasingly prevalent. The deep models have revolutionized the computer‐aided diagnosis systems in terms of performance. However, DL models are often criticized as black boxes due to their complex and opaque internal design of numerous interconnected layers and parameters, making it challenging to interpret their decision‐making process. This limitation has led to growing interest in eXplainable AI (XAI) approaches that aim to provide insights into model behavior and improve the explainability of predictions. However, integrating explainability into DL models to enhance the trustworthiness of skin lesion diagnosis systems is still in its early stages. Moreover, many methods focus mainly on subjective visualizations of explainability without achieving advanced use of these methods. This trade‐off limits the practical use of aided diagnosis systems in clinical settings. Within this framework, this study thoroughly proposes a systematic review of the literature on the latest research and advances in the integration of XAI in DL‐based skin lesion diagnosis contexts. We have systematically categorized existing XAI methods according to their application objective, namely methods designed to provide visualizations for end‐users, methods aimed at improving the model itself, and methods focused on the quantitative validation of these XAI methods. Thus, this review provides a thorough and systematic analysis of commonly used XAI methods in previous studies, identifies current challenges, motivations, and emerging recommendations to lead future research.