Domain shifts limit the generalisation of deep learning models for skin cancer detection, particularly when trained on dermoscopic images but deployed on clinical images. This study evaluates supervised and unsupervised domain adaptation techniques to improve model performance on a diverse set of clinical images. We introduce the IMPS dataset, a varied collection of clinical skin lesion images, to assess robustness under real-world conditions. Experimental results show that unsupervised methods, particularly Domain-Adversarial Neural Networks (DANN), provide better generalisation than supervised approaches. These findings suggest that evaluating models on limited datasets may give an incomplete picture of their reliability. Future research should test these approaches on additional clinical datasets that were not part of this study to better assess their suitability for real-world applications.