<![CDATA[Skin lesions are critical dermatological indicators that require early detection to prevent severe outcomes such as melanoma. Traditional Convolutional Neural Network (CNN) architectures employed for categorizing these lesions frequently encounter significant hurdles, notably disproportionate class distributions and a lack of transparency, functioning essentially as opaque "black boxes" during inferential processes. To mitigate these limitations, the current research deploys a ResNet-50 framework augmented by a Convolutional Block Attention Module (CBAM) to refine spatial and channel feature prioritization, alongside the integration of Gradient-Weighted Class Activation Mapping (Grad-CAM) to yield interpretable visualizations. The empirical analysis utilized the HAM10000 repository, incorporating a preprocessing pipeline that encompassed spatial resizing, pixel normalization, and data augmentation, subsequently trained via a bipartite transfer learning methodology. Quantitative metrics reveal that the CBAM-integrated architecture elevates the baseline global accuracy from 82.00% to 86.83%, while simultaneously augmenting the Macro F1-Score from 68.00% to 77.00%. Qualitative evaluation using Grad-CAM shows sharper and more localized heatmaps, indicating that the attention mechanism successfully guides the model to focus on clinically relevant lesion areas. These findings suggest that combining attention mechanisms with explainable AI not only enhances classification performance but also provides visual transparency, supporting clinical interpretation. This approach is expected to improve trust and reliability in automated skin lesion classification systems]]>
