A trustworthy and explainable deep learning framework for skin lesion detection in smart dermatology

• Published in: Engineering applications of artificial intelligence Vol. 159; p. 111594
• Main Authors: Eita, Mohammad A., Rizk, Hamada
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2025
• Subjects: Artificial intelligence; Computer vision; Deep learning; Dermatology; Machine learning; Object detection; Skin diseases; Skin lesion detection; Smart hospitals; You Only Look Once algorithms; Deep learning; Computer vision; Skin lesion detection; Dermatology; Object detection; Machine learning; You Only Look Once algorithms; Skin diseases; Smart hospitals; Artificial intelligence
• ISSN: 0952-1976
• DOI: 10.1016/j.engappai.2025.111594

The rapid evolution of artificial intelligence (AI) and deep learning profoundly impacts medical imaging, where it significantly enhances diagnostic accuracy. However, the effective deployment of AI systems in clinical settings, especially for skin lesion detection and diagnosis, requires not only high accuracy but also transparency and robustness to gain the trust of healthcare professionals. This is particularly crucial considering the challenges posed by varying sensor quality, lighting conditions, and lesion diversity. In this paper, we introduce a novel framework based on the You Only Look Once (YOLO) model that addresses these critical needs by enhancing both the explainability and performance of skin lesion detection models. Early and accurate identification of skin lesions is essential for the timely treatment and management of dermatological conditions. Traditional diagnostic methods, such as visual assessments by dermatologists, are often labor-intensive, subject to interpretative variability, and prone to inaccuracies, especially in cases involving atypical or subtle lesions. Our approach incorporates advanced data augmentation techniques to improve the model’s generalization capabilities across diverse clinical conditions. Additionally, we integrate saliency maps to provide visual explanations of the model’s predictions, allowing clinicians to understand the decision-making process and ensuring alignment with established clinical knowledge. Comparative analyses with the state-of-the-art models highlight the superior performance of our proposed framework, with significant improvements in the harmonic mean of precision and recall (F1-Score), and the Mean Average Precision (mAP50). The results underscore the effectiveness of our framework and how it advances the application of trustworthy AI in dermatology, paving the way for more reliable and informed clinical decisions in the diagnosis and treatment of skin conditions.