Synthetic CT images for semi-sequential detection and segmentation of lung nodules

Accurately detecting and segmenting lung nodules from CT images play a critical role in the earlier diagnosis of lung cancer and thus have attracted much interest from the research community. However, due to the irregular shapes of nodules, and the low-intensity contrast between the nodules and other lung areas, precisely segmenting nodules from lung CT images is a very challenging task. In this paper, we propose a highly effective and robust solution to this problem by innovatively utilizing the changes of nodule shapes over continuous slices (inter-slice changes) and develop a deep learning based end-to-end system. Different from the existing 2.5D or 3D methods that attempt to explore the inter-slice features, we propose to create a novel synthetic image to depict the unique changing pattern of nodules between slices in distinctive colour patterns. Based on the new synthetic images, we then adopt the deep learning based image segmentation techniques and develop a modified U-Net architecture to learn the unique color patterns formed by nodules. With our proposed approach, the detection and segmentation of nodules can be achieved simultaneously with an accuracy significantly higher than the state of the arts by 10% without introducing high computation cost. By taking advantage of inter-slice information and form the proposed synthetic image, the task of lung nodule segmentation is done more accurately and effectively.