MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Imaging Scientific Session
  5. Non-Rigid 4D-CT Image Registration Using An Unsupervised Deep Convolutional Neural Network

Non-Rigid 4D-CT Image Registration Using An Unsupervised Deep Convolutional Neural Network

Y Lei*, Y Liu , T Wang , Z Tian , S Tian , W Curran , T Liu , K Higgins , X Yang , Emory Univ, Atlanta, GA

Presentations

(Thursday, 7/18/2019) 7:30 AM - 9:30 AM

Room: 221AB

Purpose: To develop a deformable 4D-CT image registration method capable of accurately modeling respiratory motion, which can estimate tumor and organ-at-risk (OAR) motion throughout the respiratory cycle to guide treatment planning and improve the tumor targeting accuracy during lung stereotactic body radiotherapy (SBRT).

Methods: We propose an unsupervised deformable image registration framework for 4D-CT using generative adversarial networks (GANs). During the training stage, patch-based structural features between moving and fixed images were extracted using dilated convolutions, and a 3D displacement vector field (DVF) was obtained from an end-to-end fully-convolution-network (FCN). The deformed patch was obtained by applying the DVF to the moving patch. Then, a convolution neural network (CNN)-based discriminator was used to verify the accuracy of the deformed patch against with fixed patch. During the generation stage, a full DVF from any two phases of a new 4D-CT image was reconstructed through the fusion of patch-based DVF estimated by the trained network. Finally, the deformed CT image was obtained by applying the full DVF to the moving image. This registration technique was validated with a clinical study of 10 patients’ cases, each with 10 phases within a 4D-CT. We used the leave-one-out method to evaluate the proposed algorithm.

Results: The mean absolute error (MAE), peak-signal-to-noise-ratio (PSNR) and normalized cross-correlation (NCC) indexes between the fixed and deformed images were 15.6±5.3HU, 36.5±3.8dB, and 0.97±0.02. The mean target registration error (TRE) of multiple landmarks in the post-registration images was less than 2.5 mm.

Conclusion: We have proposed a novel unsupervised deep-learning-based non-rigid 4D-CT image registration method, and demonstrated its feasibility and reliability. This 4D-CT non-rigid registration technique could be a promising tool for estimating respiratory motion for accurate tumor and OAR delineation, for improving the planning target volume (PTV) coverage and OAR sparing.

Keywords

Not Applicable / None Entered.

Taxonomy

IM- CT: 4DCT

Contact Email

Additional Links