U-net is a commonly used medical image segmentation network, but it still has the disadvantages of poor generalization ability and easy over-fitting in convolutional neural networks. In view of its shortcomings, a full convolution lung nodule segmentation network is studied, adding dropout layers , using new activation function, loss function, optimizer, etc, to improve the network structure. The improved network has a higher recall rate. Then, the local linear embedding algorithm with improved reconstruction weights is used to extract features, and finally the XGBoost classifier is used for final classification. Through experimental comparison, it is confirmed that the pulmonary nodule detection combined with the above two algorithms has a higher accuracy rate.