Melanoma is the most aggressive type of skin cancer, significantly affecting human health and life expectancy. Automated melanoma classification has become an important challenge; a promising technique to address this issue is the use of a Deep Convolutional Neural Network (CNN). The success of CNN techniques highly depends on having a sufficient amount of skin lesion data to build a well-performing predictive model. Due to the sensitivity of medical data, the performance is often impacted by the overfitting problem. Effective solutions to this issue are firstly, data augmentation, which permits to generate a large amount of artificial data. Secondly, using the pre-trained VGG16 model, researchers can benefit from transfer learning, where the network’s knowledge from a large dataset can be transferred to the skin cancer classification task. On the other hand, to improve the diagnosis precision, it is important to merge multimodal information collected from various image representation. The idea is that different imaging modalities contain extensive information that are different and complementary to each other. Motivated by these issues, this paper presents a multimodal approach based on the cooperation of four imaging modalities and using transfer learning scenario. Each pre-trained VGG16 was trained by an imaging modality, the cross-modality fusion at the decision-making level was ensured by Sum and Product functions. The experiments were carried out on 170 dermoscopic images. After a data augmentation process, results show that the VGG16 trained with multimodal images performances well compared to VGG trained with single-modal images (between 0.98% and 0.99%). These results provide new evidence for the profit of multimodal imaging in medical applications.