4D-CT Deformable Image Registration Using an Unsupervised Deep Convolutional Neural Network

Four-dimensional computed tomography (4D-CT) has been used in radiation therapy to allow for tumor and organ motion tracking throughout the breathing cycle. It can provide valuable information on the shapes and trajectories of tumor and normal structures to guide treatment planning and improve the accuracy of tumor delineation. Respiration-induced abdominal tissue motion causes significant problems in effective irradiation of abdominal cancer patients. Accurate and fast deformable image registration (DIR) on 4D-CT could aid the treatment planning process in target definition, tumor tracking, organ-at-risk (OAR) sparing, and respiratory gating. However, traditional DIR methods such as optical flow and demons are iterative and generally slow especially for large 4D-CT datasets. In this paper, we present our preliminary results on using a fast-unsupervised generative adversarial network (GAN) to generate deformation vector fields (DVF) for 4D-CT DIR to help motion management and treatment planning in radiation therapy. The proposed network was trained in an unsupervised fashion without the need of ground truth DVF or anatomical labels. A dilated inception module (DIM) was integrated into the network to extract multi-scale motion features for robust feature learning. The network was trained and tested on 15 patients’ 4D-CT abdominal datasets using five-fold out cross validation. The experimental results demonstrated that the proposed method could attain an accurate DIR between any two 4D-CT phases within one minute.