4D CT Deformable Registration Using Unsupervised Deep Learning for Lung Stereotactic Body Radiation Therapy

Abstract

In current lung stereotactic body radiation therapy (SBRT), physicians assess the tumor motion on the movie loops of the 4D CT images to select treatment phases for gated lung SBRT. However, physician’s manual selection of the treatment phase is time-consuming and subjective. To address this challenge, we propose to develop a 4D CT deformable registration method, namely LungRegNet, to accurately assess the respiratory motion of lung/tumor using 4D CT images which could help choose the optimal treatment phases for lung SBRT. The LungRegNet uses deep learning and consists of two sub-networks which are CoarseNet and FineNet. As the name suggests, CoarseNet predicts large lung motion on a coarse scale image while FineNet predicts local lung motion on a fine scale image. Both the CoarseNet and FineNet include a generator and a discriminator. The generator is trained to directly predict the displacement vector field (DVF) to deform the moving image. The discriminator is used to introduce additional DVF regularization to encourage realistic motion prediction by distinguishing the deformed images from the original images. CoarseNet is first trained to deform the moving images. The deformed images are then used by the FineNet for FineNet training. To increase the registration accuracy of the LungRegNet, we generate vessel-enhanced images by generating pulmonary vasculature probability maps prior to the network prediction. We performed five-fold cross validation on 10 lung 4D CT datasets, in which 10 CT images were acquired during a whole respiratory period for each patient. In addition, we tested our method using 10 datasets from public database DIRLAB that provided 300 manual landmark pairs per case for target registration error (TRE) calculation. The mean and standard deviation of TRE was 1.00±0.53mm for our patients and 1.59±1.58mm for the external DIRLAB datasets. The results indicated that our LungRegNet achieved better registration accuracy in terms of TRE than other available state-of-the-art deep learning-based methods on DIRLAB datasets. We have developed a novel unsupervised deep learning-based method to rapidly register 4D CT lung images and demonstrated its clinical feasibility and accuracy. This LungRegNet could accurately estimate patient-specific respiratory motion for lung SBRT to help physicians objectively find the optimal treatment phases, and enable more accurate accumulative dose calculation and thus more accurate plan evaluation. This tool will be further tested in prospective trials to determine if it could improve lung tumor delineation, reduce dose delivered to the organs-at-risk, and potentially improve survival of many lung cancer patients.