Accurate classification of pulmonary nodules is critical due to their strong association to lung cancer. However, despite advances in CT imaging, the inherent variability and morphological ambiguity of nodules remain significant barriers to reliable differentiation between benign and malignant forms. To overcome this challenge, we propose a novel multi-scale classification framework for pulmonary nodules based on structure-aware learning. The model introduces directionally-constrained von Mises-Fisher (vMF) clustering and spatial assignment coefficients to learn discriminative structural prototypes in an unsupervised manner. These prototypes effectively capture stable and interpretable patterns within pulmonary nodules. Moreover, the model integrates multi-scale features from the first four ResNet-18 stages via a Feature Pyramid Network (FPN), enabling the construction of a unified high-dimensional representation that encodes both fine-grained details and contextual semantics. CBAM attention modules are incorporated at each backbone stage to enhance structural responsiveness. The final classification is performed using a multilayer perceptron (MLP) based on a probabilistic visual concept embedding derived from the fused features. Evaluated on the LIDC-IDRI dataset, MAG-VCNet achieves a test AUC of 90.57%, outperforming several state-of-the-art methods. These results demonstrate the effectiveness of combining structure-aware learning, vMF-based visual concepts, and multi-scale fusion for robust and interpretable pulmonary nodule classification.