Fuzzy cognitive explainable AI framework integrating ResNet-50 and causal clinical concept reasoning for skin lesion...

Melanoma is one of the most aggressive forms of skin cancer, and its early detection from dermoscopic images remains a significant clinical challenge. Deep learning (DL) convolutional neural networks (CNNs) have achieved remarkable results in medical image analysis. However, their inherent opacity is a barrier to trustworthy clinical adoption. Clinicians need not only accurate predictions but also clear, understandable reasons behind those predictions. Although post-hoc explainability techniques such as Grad-CAM and SHapley Additive exPlanations (SHAP) are available, but there is a critical gap. They produce low-level, pixel-based saliency maps that do not connect meaningfully to the clinical concepts dermatologists actually use, namely the ABCD criteria (Asymmetry, Border, Colour, and Diameter). No existing framework captures the causal relationships between these concepts, nor quantifies the uncertainty that arises when the predictive. The model and its explanation disagree. This work addresses that gap through a two-stage hybrid framework. Fuzzy-Cognitive Explainable AI (FC-XAI). In the first stage, a ResNet-50 model is trained for melanoma detection. In the second stage, we employ SHAP to identify salient image regions and map them to high-level clinical concepts. These concepts form the nodes of a Fuzzy Cognitive Map (FCM), a dynamic graph-based model that captures structured concept-level reasoning. The FC-XAI framework was evaluated on the ISIC melanoma dataset. The base DL model achieved an AUC of 0.9832 (95% CI: [0.90, 1.00]), an accuracy of 84%, and a recall of 80% for the malignant class. The generated FCMs produced clear, structured concept-level explanations (e.g., high Asymmetry and Border irregularity strongly activate Malignancy). The uncertainty metric reveals that medium-uncertainty scenarios achieved 82.4% accuracy compared with 57.1% for low-uncertainty scenarios, a finding with direct triage indications. External experimental validation on the PH2 dataset yielded an AUC of 0.9498 with clinically coherent FCM descriptions, supporting reasonable cross-dataset generalisation. By adding a predictive Deep Learning model with a transparent structured concept-level causal model, FC-XAI represents a concrete step towards clinically reliable AI system in dermatology.