AI systems outperform dermatologists in diagnosing pigmented lesions in pigmented lesions

A research team from Yamamoto Yuko, Nakajima Karako, a department of dermatology at Kindia University and other institutions has collaborated with the Department of Engineering at Kindai University and other institutions to develop a diagnostic system that uses artificial intelligence (AI) to accurately identify the types of treatments that support. A paper on this study has been published online CyreusInternational Medical Journal on June 5th, 2025.

1. Key Points

• We demonstrated superior diagnostic accuracy compared to dermatologists when identifying five types of lesions, confirming the usefulness of the system in early detection and treatment decisions.
• Using AI, we have developed a system that accurately classifies and supports the diagnosis of five difficult-to-diagnostic pigment lesions.
• It helps in establishing ways to accurately identify pigmented lesions, reduce the risk of abuse, and support appropriate treatment decisions.

2. Research background

Pigment lesions come in many different types, such as liver, ephelide, and acquired cutaneous melanocytosis, sun lentigo, and lentigomarigna melanoma, but are often visually similar, making discriminatory diagnosis difficult. On the other hand, proper treatment for these lesions varies greatly depending on the type, and this directly affects the feasibility and choice of laser treatment, so accurate diagnosis is essential. For example, inappropriate laser use can exacerbate melasma, and delaying the surgical resection required for Lentigo Maligna melanoma due to misdiagnosis can have serious consequences. In recent years, imaging diagnostic techniques using deep learning models have attracted attention, and the results of these studies have shown that they are comparable or better than the accuracy of dermatologists to distinguish skin lesions. Deep learning-based imaging has successfully detected melanoma, but the development of diagnostic support systems is required due to insufficient research on benign and malignant pigmented lesions directly related to laser treatment planning.

3. content

The researchers developed a classification system using deep learning models InceptionResnetv2 and Densenet121 to identify five types of complexion lesions: melasma, ephelide, acquired dermal melanotosis, sun lentigo, and Lentigo maligna/lentigo maligna melanoma. Training and validation were performed using 432 clinical images, and their diagnostic accuracy was compared to the diagnosis of nine board-certified dermatologists (experts) and 11 uncertified dermatologists (non-experts). Both models showed diagnostic accuracy of 87% and 86%, respectively. Both models performed median diagnostic accuracy of 80% for experts and 63% for non-experts. In particular, when identifying LM/LMM, both models achieve 100% sensitivity, suggesting their potential use as a diagnostic support tool in clinical practice.

Based on these results, the developed deep learning model far outweighs the dermatologist's accuracy in diagnosing pigment lesions, and may contribute to diagnostic support and decisions on appropriate treatment plans.