Lung cancer stands as the most lethal among malignant cancers globally, underscoring the urgency of early detection for effective treatment. The reliance on the experience of radiologists for accurate pulmonary nodule diagnosis, coupled with their substantial workload, prompted the development of Artificial Intelligence (AI)-based Computer-Aided Diagnosis (CAD) systems. While these systems have alleviated the burden on radiologists and improved diagnosis accuracy, concerns about model reliability and interpretability hinder widespread clinical application. This paper introduces an AI technique which is an attention-based deep learning architecture designed for lung nodule diagnosis. It not only predicts malignancy in nodule lesion but also identifies its characteristics, providing visual interpretability. Leveraging a 3D U-Net for probability maps of lung nodules, an anatomical attention module integrates spatial attention to informative details into the classification network. Fine-grained activation maps are generated by the soft activation map module to visualize nodule characteristics. It demonstrates significant performance with a test AUC of 0.93 on the publicly available dataset of LIDC containing 624 participants. Importantly, the incorporation of characteristic identification tasks enhances the accuracy of malignancy classification, as validated by experimental results. This work advances the AI state of the art in pulmonary nodule diagnosis by introducing characteristics. It is enhancing the nodule classification accuracy with accuracy of 0.956, precision of 0.892, specificity of 0.951, and sensitivity of 0.989.