Lung cancer (LC) is the second most common cancer worldwide. Lung nodules are the predominant biomarker for LC diagnosis and treatment. Today, expert radiologists conduct nodule analysis, considering texture, size, and nodule morphology. Also, radiologists exploit contextual information related to nodule localization and the configuration of the lung structures. Computational strategies, however, have been principally dedicated to isolated nodule analysis, without complementary contextual lung information. This study introduces a transformer-based strategy that integrates both 3D nodule scale observations and lung-scale context to enhance malignancy classification. Lung context is here codified from a standard transformer encoder, providing a robust contextual embedding vector. In parallel, a specialized multi-head attention encoder captures the nodule scale local information to support malignancy classification, which is thereafter integrated with contextual information. The proposed approach was validated on the LIDC-IDRI dataset, achieving a 96.44% and 95.53% of AUC and recall, respectively, outperforming current state-of-the-art approaches.