Pulmonary nodule detection plays an important role in lung cancer screening with low-dose computed tomography (CT) scans. Although promising performance has been achieved by deep learning based nodule detection methods, it remains challenging to build nodule detection networks with good generalization performance due to unbalanced positive and negative samples. In order to overcome this problem and further improve state-of-the-art region proposal network methods, we develop a novel deep 3D convolutional neural network with an Encoder-Decoder structure for pulmonary nodule detection. Particularly, we utilize a dynamically scaled cross entropy loss to reduce the false positive rate and compensate the significant data imbalance problem. We adopt the squeeze-and-excitation structure to learn effective image features and fully utilize channel inter-dependency. We have validated our method based on publicly available CT scans from LIDC/IDRI dataset and its subset LUNA16 with thinner slices. Ablation studies and experimental results have demonstrated that our method could outperform state-of-the-art nodule detection methods by a large margin, with an average FROC score of 86.2% on LUNA16, and an average FROC score of 77.3% on LIDC/IDRI dataset when trained on LUNA16 only.