New AI system automates classification of skin lesions

Skin cancer is a global health challenge, with nearly 1.5 million new cases diagnosed in 2024 according to World Health Organization (WHO) data. In response, the University of Alicante (UA) and the Sant Joan d'Alacant University Hospital have collaborated to launch MEL-IA (MobilE skin Lesion dIAgnosis), an artificial intelligence system designed to automate the classification of skin lesions.

Early detection remains vital to improving clinical outcomes for skin cancer. This technology supports healthcare staff in evaluating lesions while seamlessly integrating into the hospital's existing IT infrastructure.

The system incorporates a mobile app to capture images of skin lesions—such as moles or spots—and record clinical data, an AI model to perform the classification, and complex integration models that securely connect all data with hospital systems.

The authors of this study, published in the scientific journal Journal of Medical Systems, are Alberto de Ramón, Daniel Ruiz, and Marcelo Saval, lecturers in the Department of Computer Technology and Computation at the UA, together with Pablo Candela, all members of the Bio-inspired Engineering and Health Informatics (IBIS) research group at the UA. The IBIS group has spent over a decade researching clinical decision-support systems for diagnosing skin lesions. Representing the Sant Joan d'Alacant University Hospital, José María Salinas and Diego Guijarro from the IT Service co-authored the work.

Beyond developing the classification algorithm, the team has delivered a comprehensive technology package capable of embedding directly into a hospital's IT network. The system links image capture, AI analysis, secure storage, and the confidential exchange of clinical data.

Skin lesions and diagnostic performance

Unlike tools designed solely to distinguish between malignant and benign lesions, MEL-IA provides differential categorization across five major skin lesion types: melanoma, naevus, basal cell carcinoma, actinic keratosis, and benign keratosis.

To train and validate the model, researchers integrated over 15,000 dermatoscopic images alongside clinical patient data, including age, sex, and anatomical lesion location, to enhance classification accuracy.

Results indicate that the system achieves an overall accuracy of 86%. The UA researchers noted that for melanoma detection, MEL-IA achieved a sensitivity of 88%, while reaching 92% for basal cell carcinoma, adding that the highest overall scores corresponded to naevi and basal cell carcinomas.

A key highlight of the project is that the technology moved beyond the experimental stage to deployment at the Sant Joan d'Alacant University Hospital to evaluate performance in a live healthcare environment. The team highlighted that they processed 980 dermatological studies with response times under one second.

Furthermore, the system maintains a longitudinal record of lesions, embedding images, diagnoses, and medical observations over time.

The researchers emphasize that MEL-IA serves as a clinical decision-support tool rather than an autonomous diagnostic system. Key next steps include conducting prospective studies with healthcare professionals, testing direct image acquisition via smartphones, and expanding the range of lesion types the system can classify.