Purpose: Four dimensional (4D) cone beam computed tomography (CBCT) image quality is degraded by insufficient projection of each respiration phases, which can be solved by simultaneous motion estimation and image reconstruction (SMEIR) method by iteratively performing motion-compensated images reconstruction and motion estimation. However, the motion estimation step used projection intensity matching method leading to deformation vector fields (DFVs) having less accuracy inside of lung compared to lung boundary region. We developed a convolutional neural network (CNN) based approach to improve the inside lung DVFs accuracy and decrease the computational time.
Methods: We built two U-net based architectures to derive DVFs inside of lung from lung boundary DVFs. For the first architecture (U-net-3C), each direction of DVFs from SMEIR method obtained from motion-estimation step served as one input channel. The output contains 3 channels denoting each direction of the high-quality DVFs. For the second the architecture (U-net-4C), a CT image channel served as an additional input channel to combining heterogeneous properties of inside lung with inside lung motion to improve DVFs accuracy. This model is evaluated on 85 DVFs of 11 patients using a 5-fold cross validation. Displacements of manually tracked land-marks are extracted to compare with predicted DVFs to evaluate the accuracy of predicted DVFs.
Results: Average magnitude of displacement 3D vector between U-net-3C predicted DVFs and manually-tracked landmarks is 3.34 mm and that between U-net-4C predicted DVFs and manually-tracked landmarks is 3.02 mm, while the difference is 5.04 mm for the original SMEIR. Once the model training process terminated, it takes around 10 second to obtain updated DVFs.
Conclusion: In this work, a U-net based approach that deriving DVFs inside lung from lung boundary to improve the accuracy of DVFs generated from 2D-3D-registration and decrease the computational time has been proposed.
Not Applicable / None Entered.