Click here to

Session: CT and CBCT: New Technologies, Algorithms, and Emerging Applications II [Return to Session]

Accurate and Efficient Deep Neural Network Based Deformable Image Registration Method in Lung Cancer

Y Ding1*, Z Liu2, H Feng3, J Holmes3, Y Yang3, N Yu3, T Sio3, S Schild3, B Li1, W Liu3, (1) Arizona State University,Tempe, AZ, (2) University of Georgia, Athena, Georgia,(3) Mayo Clinic, Phoenix, AZ

Presentations

SU-F-201-5 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]

Room 201

Purpose: Deformable image registration (DIR) is an essential technique required in many applications of radiation oncology such as contour propagation and dose deformation. However, conventional demons-based and/or B-spline-based DIR approaches typically take minutes to register one pair of 3D CT images and the resulted deformable vector fields (DVFs) are specific to the pair of images generated from. Thus, it is less appealing in clinic use. In this work, we propose a deep learning-based approach to achieve fast and accurate DIR for 3D CT images in lung cancer.

Methods: We used a deep neural network called VoxelMorph to build a model to generate DVFs based on a 3D lung CT images dataset composed of 104 lung cancer patients. In total, 170 pairs (initial CTs and verification CTs) of lung CTs were used for training and 9 pairs of lung CTs were used for testing. The verification CTs usually were taken a couple of weeks after the initial CTs. The performance of the proposed method was evaluated by measuring the similarity between the ground truth CTs (initial CTs) and the synthetic CTs, which were generated by registering the repeated CTs to the initial CTs via the DVFs generated by the trained model. CT-number-absolute-difference volume histogram (CDVH) and mean absolute error (MAE) were used as the evaluation metrics. The times to generate the synthetic CTs were also recorded.

Results: The proposed approach achieved a speed of 183.8± 0.4 ms and a MAE of 68.3±39.4 HU for the testing set. CDVH of a typical patient showed that only 5% of the voxels have a CT-number-absolute-difference larger than 168 HU.

Conclusion: A deep neural network-based DIR approach was accurate and efficient and can be used for registering any lung cancer CTs once the model was trained.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Share: