Real Time 4D-Cone Beam CT Accurate Estimation via Convolutional LSTM Network

Recently the constructing Cone Beam CT(CBCT) methods based on few 2D projection images for determining anatomical information have gained momentum and drawn more interest. Although high quality 4D-CBCT is also capable to monitor accurate moving targets before Radiation Therapy (RT) beam on, the iterative reconstruction time cost is expensive. The existed deep learning based 4D-CBCT motion estimation scheme usually relied on conventional architectures such as Convolutional Neural Network (CNN). But these methods only work in a single phase, ignoring the temporal information of 4D-CBCT. In this study, we first introduced the concept of spatio-temporal combination in the field of CBCT reconstruction,and proposed a cascade model composed of Convolutional LSTM (ConvLSTM) and Principal Component Analysis (PCA) model with prior knowledge to map a single 2DX-ray to 3D-CBCT (e.g. one phase of the 4D-CBCT) for achieving real time accurate 4D-CBCT. PCA coefficients were randomly sampled to generate Digital Reconstruction Radiographs (DRRs) for training the cascade network. The noise-free DRRs were calibrated using data enhancement to simulate the scattering and quantum noise. The weight of the PCA coefficients was introduced into the loss function to ensure a more effective training result. The experimental results showed that the proposed method achieved the best performance compared with the other state-of-the-art methods. This study demonstrated the spatio-temporal combination method supplied a promising solution for accurate real time 4D-CBCT prediction.