(Tuesday, 7/31/2018) 4:30 PM - 6:00 PM
Room: Room 202
Purpose: To explore the potential of convolutional neural network (CNN) for 4D-CT lung deformable image registration (DIR).
Methods: The proposed CNN takes a pair of target and moving images as input and predicts the voxel-by-voxel dense deformation vector field (DVF) directly. Eight of the 10 DIRLAB lung datasets were used to train the network. The remaining two of the datasets were used to test the trained network. Only the End-Inhalation (EI) and End-Exhalation (EE) phases of a respiratory cycle of the DIRLAB 4D-CT lung datasets were used. Artificial DVFs were created to increase the number of training datasets. The network was trained using both supervised and unsupervised training. The supervised training minimizes the difference between the ground truth DVFs and the predicted DVFs. The unsupervised training maximize the cross-correlation between the target images and the deformed moving images. The CNN-predicted DVFs were filtered using Gaussian smoothing to get the final predicted DVFs. Target Registration Errors (TREs) of the registration results were calculated and compared with conventional Horn-Schunck (HS) optical flow methods for evaluation.
Results: The proposed CNN-based DIR could predict voxel-by-voxel DVF directly from the EI and EE phases of a respiratory cycle. The network trained using the combined supervised and unsupervised training performed better than the network trained using either the supervised or unsupervised training. The registration time for image sizes of 128*128*64 was around 2 seconds on a GeForce GTX 1080 Ti GPU. The average TRE of the proposed method was 3.38 mm which was better than the single-scale HS (6.62 mm) and worse than the multi-scale HS (2.16 mm) for the two testing cases.
Conclusion: The proposed CNN-based deformable image registration method has the potential to perform equally well as or better than the conventional HS methods for 4D-CT lung DIR especially for large deformation DIR.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
