Achieving the reliable and automatic pulmonary nodule malignancy prediction is a critical step in reducing the workload of physicians and advancing towards intelligent medical treatment. The ‘black-box’ of deep neural networks is a major obstacle in achieving reliable prediction. An interpretable neural network framework for pulmonary nodule malignancy prediction is proposed to ensure reliable predictions. It mainly includes image generation module, mapping network module, reversible network module and reliable classifier module. The latent space constructed based on the image generation module can completely and accurately represent the features of pulmonary nodules. By mapping the input images to the latent space, the accurate representations of pulmonary nodules can be obtained, and then they are mapped to the feature space through the reversible network module. Finally, based on data-driven knowledge in feature space and knowledge-driven medical knowledge, the reliable malignancy prediction is achieved. The experiment shows that our model can accurately extract the features of pulmonary nodules and reliably predict malignancy based on those specific features and medical knowledge. In addition, this model can be used to analyze the difference between data-driven knowledge and knowledge-driven knowledge, which helps enrich and refine medical knowledge while interpreting the basis for prediction.