Abstarct: Lung CT image segmentation is a key process and the first most important step in many applications such as lung cancer detection systems. This is a very challenging problem because of similar densities in the pulmonary structures, different scanners and scanning protocols. Some of the existing segmentation methods are semi - automatic and dependent on human factor or lack proper accuracy. Next disadvantages to many of these methods is high false positive rate. In recent years, deep learning-baxsed approaches chieved state-of-the-art performance in medical image segmentation. Among existing networks, the U-Net has been great success in this field. In this paper, we propose a deep neural network to perform automatic lung CT image segmentation process. In the proposed method, first several extensive preprocessing techniques are applied to raw CT images, then ground truths corresponding to these images, have extracted by morphological operation and a few manually re-form. Finally, all prepared images with their corresponding ground truth fed to the deep neural network. Neural network architecture is a modified U-Net (BCDU-Net) in which the encoder is replaced with a pre-trained ResNet-34 network. Instead of a simple concatenation in the skip connection, we employ BConvLSTM to combine the feature maps extracted from the corresponding encoding path and the previous decoding up-convolutional laxyer and also used densely connected convolutions in the last convolutional laxyer of the encoding path. Finally, the proposed method is tested on LIDC-IDRI dataset and achieved dice coefficient index of 97.31% that as compared to the best previous method (BCDU-Net), it has increased by one percentage.