Catheters, lines, and tubes are life-supporting devices often inserted into patients undergoing medical procedures. Radiologists examine chest X-ray images immediately after insertion to avoid serious complications from misplaced devices. We seek to develop a machine learned automatic evaluation of these inserted devices to provide faster interpretation and improved patient outcomes. A previous approach to address this challenge with machine learning synthesized lines on radiographs to generate annotated data, then used a U-Net style segmentation network. Separate studies have demonstrated the value of adding residual connections to Fully Convolutional Networks (FCN) used for biomedical image segmentation. Our approach extends this research by incorporating a ResNet50 backbone into U- Net and by proposing a novel transfer learning scheme. First, a ResNet50 model pre-trained on the ImageNet data set is fine tuned for an auxiliary classification task. Second, the full segmentation network, using the ResNet50 backbone, is trained on the primary task performing semantic segmentation of lines. Our approach was developed and tested with two data sets: 1) the publicly available Stanford CheXpert data set, and 2) an IRB-approved set of pediatric radiographs, both of which we manually annotated with the line locations. We demonstrate that this transfer learning scheme outperforms learning from scratch and leads to a viable approach for rapid line evaluation in chest radiography.