2141 - AI-Based Generation of ITV for 4DCT Data
Presenter(s)
K. Li1, C. Liu2, A. N. De Leo2, and S. Samant1; 1University of Florida, Gainesville, FL, 2Department of Radiation Oncology, University of Florida College of Medicine, Gainesville, FL
Purpose/Objective(s): To generate internal target volume (ITV) of lung tumors on 4DCT using a hybrid VoxelMorph neural network combined with optical flow.
Materials/Methods: To evaluate the performance of our hybrid VoxelMorph model versus conventional VoxelMorph, the study used 15 clinical 4DCT image sets (1.7x1.7x3mm3 voxels):7 assigned to training, 2 assigned to validation, and 6 used for testing. The hybrid model was trained to learn the spatio-temporal transformation between two breathing phases, involving conventional VoxelMorph for spatial deformation; and an additional Farneback model for computing temporal optical flow. Loss function comprised sum of normalized local cross-correlation, displacement vector regularization term, and dice score between warped and ground-truth gross tumor volume (GTV). For the testing stage, image and GTV (contours drawn by user) in phase 0 are selected as fixed image, and were input to the model together with moving image sets in phases 1 to 9. The prediction ITV was then generated by expanding GTVs into all 10 phases. The generated ITV was compared with ground truth ITV delineated by in-house oncologists. Dice score (DSC), 95-percentile Hausdorff distance (HD95), and mean surface distance (MSD) were computed. Computations were carried out using one A100 GPU and four ROME CPUs.
Results: ITV DSC for model with and without optical flow as 0.85 ± 0.05 and 0.79 ± 0.03, respectively with significant statistical difference (p<0.05). ITV HD95% for model with and without optical flow was 3.07 ± 0.11 mm and 4.76 ± 1.60 mm, respectively with no significant difference (p>0.05). The ITV MSD of model with and without optical flow is 1.04 ± 0.18 mm and 1.47 ± 0.44 mm, respectively with no significant difference (p>0.05), indicating good equivalency between ITVs. Model computation time per patient with optical flow is 40 secs.
Conclusion: Our hybrid VoxelMorph outperformed conventional VoxelMorph in predicting tumor ITV based on 4DCT in terms of DSC, HD95% and MSD user drawn GTV contour set on single phase of 4DCT in 40 secs and potentially could be used to assist radiation oncologists to improve workflow.