Reconstructing 3d lung shape from a single 2d image during the deaeration deformation process using model-based data augmentation

Abstract

Three-dimensional (3D) shape reconstruction is particularly important for computer assisted medical systems, especially in the case of lung surgeries, where large deaeration deformation occurs. Recently, 3D reconstruction methods based on machine learning techniques have achieved considerable success in computer vision. However, it is difficult to apply these approaches to the medical field, because the collection of a massive amount of clinic data for training is impractical. To solve this problem, this paper proposes a novel 3D shape reconstruction method that adopts both data augmentation techniques and convolutional neural networks. In the proposed method, a deformable statistical model of the 3D lungs is designed to augment various training data. As the experimental results demonstrate, even with a small database, the proposed method can realize 3D shape reconstruction for lungs during a deaeration deformation process from only one captured 2D image. Moreover, the proposed data augmentation technique can also be used in other fields where the training data are insufficient.