Room: Track 2
To verify the capability of a novel deep learning-based model, cycle consistent VoxelMorph, on Lung CT deformable image registration (DIR) and the feasibility of real-time ultra-quality 4D CT synthesis with it.
42 4D Lung CT scans without apparent motion artifacts from The Cancer Imaging Archive (TCIA) were used for training, and 3 TCIA scans from 3 different patients were used for internal validation. Another 10 scans from DIR-Lab were used for external validation. The images were isotopically resampled and cut into patches of 64x64x64. During training, CTs from different time phases of one scan were paired and fed to modified VoxelMorph, and the output displacement vector fields (DVFs) and warped volumes were penalized by smoothness and similarity. To evaluate the model and synthesize 4D images, the end of inhalation (EOI) phase of each test scans was registered to all following time phases, and the DVFs were applied to the EOI phase to generated synthesized images at different time phases. Landmarks were identified both on the true 4D CT and synthesized 4D CT to evaluate registration error.
After registration, the mean square error (MSE) and cross-correlation (CC) of TCIA scans changed from 380±210 to 139±77 and 0.93±0.04 to 0.98±0.01, respectively. As external validation, the MSE and CC of DIR-Lab scans changed from 1210±750 to 480±210 and 0.98±0.01 to 0.996±0.002, respectively. 4D CT synthesis took about 10 seconds, and primary movement including breathing and heart pumping can be observed on it. The registration error on identified landmarks was 0.6±0.4 mm.
This study verified the DIR performance of the novel cycle consistent deep learning model and showed the feasibility of 4D image synthesis with it. The model in the future will be further validated on regions other than lungs and other imaging modalities like low-quality cone-beam CT and MRI.