Nowadays, malignant tumors, especially lung cancer, have become the number one killer of people's lives. The early clinical manifestations of lung cancer are mostly isolated pulmonary nodules, and the CT images show a round white area. In this paper, we studied about the image segmentation and lung nodule judgment in the diagnosis of CT image. Firstly, a segmentation method of lung region of interest was used to separate the ROI region, and then based on the the classical convolutional neural network VGGNet network model structure, the last three layers of the full connection layer are improved, and the convolution layer is used instead of the full connection layer to maintain the accuracy of the model and accelerate the training speed of the network.