Lung cancer is the main cause of cancer-related deaths. Pulmonary nodules are the principal disease indicator, whose malignancy is mainly related with textural and geometrical patterns. Different computational alternatives have been proposed so far in the literature to support lung nodule characterization, however, they remain limited to properly capture the geometrical signatures that discriminate between each malignant class. This work introduces a multi-scale self-attention (MSA) network that accurately recovers geometrical and textural nodule maps. At each hierarchical level is recovered a set of saliency nodule maps that find non-local nodule correlations, properly representing radiological finding patterns. Validation was performed on the LICD-IDRI dataset, obtaining classification percentages that outperform the state of the art: 95.56% in accuracy, and 98.67% in AUC.