In this paper, a novel attention-based 3D-CNN (A3C) model is proposed to realize the accurate diagnosis of the radiation pneumonitis (RP), checkpoint inhibitor-related pneumonitis (CIP) and pulmonary lymphangitic carcinomatosis (PLC) from computed tomography (CT) images. In particular, the 3D convolution operators are employed due to the intrinsic ability of capturing spatial-temporal dependencies within volumetric medical images, and the channel attention mechanism is adopted to effectively address the issues of feature inter-dependencies and heterogeneity inherent in multi-class pulmonary pathology classification tasks. Moreover, by introducing multiple data augmentation techniques, the model ability of learning invariant and discriminative features from the limited annotated data is effectively enhanced. In pulmonary lesions diagnosis task, the proposed A3C outperforms the baseline models by yielding the highest accuracy of 1.0, showing the great competitiveness in providing reliable diagnostic references.