Accurate pulmonary nodule classification is vital for lung cancer diagnosis, but there are many challenges. Firstly, Self-supervised learning is widely employed in nodule classification, but position labels generated by uniformly dividing images do not offer discrimination information. Secondly, while 3D models can learn from multiple 2D slices, channel attention that captures key features from 2D to 3D overlooks characteristics across different orientations. In this paper, we propose PMNet, a nodule classification model, consisting of PSS and MWA. In PSS, we design position labels based on distribution differences of nodules to guide classification. In MWA, we introduce multiway attention to capture features from sagittal, vertical, and coronal axes to enhance distinguishing features. Moreover, we incorporate gradient boosting decision trees (GBDT) to combine shallow and deep features, which further improves accuracy. Experimental results on the LIDC-IDRI dataset demonstrate that PMNet outperforms state-of-the-art method under identical conditions, achieving an accuracy of 96.72%.