Lung cancer is the main cause of mortality from cancer globally. It is critical in clinical practise to correctly diagnose the malignancy of suspected lung nodules. However, the pathologically established lung nodule dataset is still substantially restricted and significantly uneven in benign and malignant distributions. In this article, we introduced a Semi-supervised Deep Transfer Learning (SDTL) system for distinguishing benign from malignant lung nodules. First, we apply a transfer learning technique with a pre-trained classification network to distinguish lung nodules from nodule-like tissues. Second, since the number of pathologically proved samples is limited, an iterated feature-matching-based semi-supervised technique is developed to take use of a large accessible dataset with no pathological findings. To iteratively enhance the classification network, a similarity metric function is used in the network semantic representation space to progressively include a small fraction of examples with no abnormal outcomes. In this investigation, 3,038 pulmonary nodules with pathologically proved benign or malignant labels (from 2,853 people) and 14,735 unlabeled lesions (from 4,391 participants) were gathered retrospectively. Our proposed SDTL framework exhibits higher diagnostic performance in the main dataset, with accuracy = 88.3%, AUC = 91.0%, and accuracy = 74.5%, AUC = 79.5% in the independent testing dataset. Furthermore, an ablation research reveals that using transfer learning improves accuracy by 2%, while using semi-supervised learning improves accuracy by 2.9%. The findings suggest that our proposed classification network might be an excellent diagnostic tool for suspected lung nodules and could have a potential clinical use.