Abstract
<title>Abstract</title> <p>Skin cancer remains one of the most common malignancies worldwide, with melanoma representing its most aggressiveand dangerous form. Early and reliable detection is critical for improving patient outcomes and supporting timely clinicaldecision-making. Although deep learning methods have shown encouraging effectiveness in dermoscopic image analysis,relying solely on visual information may not fully reflect the complexity of melanoma diagnosis. To address this limitation, wepropose a neuro-symbolic approach for melanoma classification that combines deep learning-based dermoscopic attributeprediction with fuzzy rule-based reasoning. A unified dataset is first constructed by harmonizing the PH2 and Derm7pt datasets,ensuring a consistent representation of shared dermoscopic attributes. Diagnostic rules are automatically extracted fromthe harmonized attribute annotations to capture explicit relationships between clinical criteria and melanoma classification.Within the proposed framework, an EfficientNet-B3 model estimates membership degrees for these attributes from lesionimages. Then, the rules are applied to the estimated degrees using fuzzy operators, allowing partial attribute expressions toguide the melanoma/non-melanoma decision, while maintaining interpretability. The best configuration, based on multimodaldeep learning and fuzzy reasoning, demonstrated an improved classification effectiveness over the image-only deep learningapproach, highlighting the benefit of integrating data-driven learning with rule-based reasoning.</p>