Dermatologists are starting to make use of Computer-Aided Diagnosis based on deep learning algorithms, which can provide them with an objective judgement during evaluation of equivocal lesions. DL algorithms can be trained to classify skin lesions with datasets of diverse nature like traditional RGB, clinical and dermoscopic images, or more experimentally, with images from other modalities, such as multispectral imaging. In this work, we have evaluated and customized the different DL approaches that exist in the state of the art to classify a dataset of +500 images acquired on skin lesions. The images were acquired with a staring multispectral imaging prototype in the visible and near-infrared ranges. The best results were obtained for a customized model VGG-16 that combined 3D convolutional layers, 3D maxpooling layers and dropout regularization, leading to an overall accuracy of 71%.