Room: Exhibit Hall | Forum 5
Purpose: To investigate the feasibility of using Convolutional Neural Network (CNN) to register phase-to-phase deformable vector field (DVF) of lung 4D-CT/4D-CBCT for motion modelling, contour propagation, 4D dose accumulation or target verification.
Methods: A CNN based deep learning method was built to directly register the deformable vector fields between the individual phase images of patient 4D-CT or 4D-CBCT. The input consists of image patch pairs while the output is the corresponding DVF that registers the image pairs. The centres of the patch pairs were uniformly across the lung and the size of the patches was chosen to cover the majority movement of deformation vectors. The network was trained to generate DVF that matches with the reference DVF generated by VelocityAI. The CNN consists of four convolutional layers, two average pooling layers and two fully connected layers. The loss function was half mean squared error. 11 sets of 4D image volumes from 9 patients with lung cancer were used for training, and the trained CNN was tested using both intra-patient (intra- & inter-fraction) and inter-patient data.
Results: Deformed images were generated based on the Velocity DVF and the CNN DVF, respectively. Main anatomical features such as diaphragm and main vessels matched well between these two images. In the diaphragm region, the cross correlation between deformed images registered by CNN and Velocity was above 0.9 for all the intra-patient cases, and above 0.87 for all the inter-patient cases.
Conclusion: CNN based deep learning achieved comparable deformable registration accuracy as Velocity. Compared to Velocity registration, the deep learning method is faster and fully automatic without user dependency, which makes it more preferable in clinical applications.