Skin cancer is still one of the most common cancers worldwide, and the survival of patients is highly dependent on a timely and accurate diagnosis [14]. Subjective, expert-dependent and time-consuming conventional diagnostic methods have inspired automated AI-supported approaches. Although Convolutional Neural Networks (CNNs) have yielded promising results for skin lesion classification, their pooling layers prevent to some extent spatial hierarchies and are suboptimal in discriminating between morphologically similar lesions. In this work, we use the HAM10000 dermatoscopic image dataset and introduce a new state of the art deep learning architecture called CapSkinNet, that is in the vein of Capsule Networks (CapsNet). In contrast to CNNs, the CapsNet employs a dynamic routing algorithm to main spatial information, so as to improve the representation ability and inter-class difference. The results reached by the CapSkinNet presented achieve a classification accuracy of 96.8% with class-wise precision of 93.8% to 96.5%, recall of 93.4% to 95.8%, and F1) also) scores between 93.9% and 96.4%. In particular, the model provides F1-scores of 96.0% for Melanocytic Nevi and 96.4% for Actinic Keratoses (AKIEC). These results demonstrate the competitive performance of CapsNet in addressing class imbalance and inter-class ambiguity. The CapSkinNet model performs better than traditional CNN baselines and holds promise as a clinically usable decision support system in dermatology.