Author: Shinichiro Mori, Isabella Pfeiffer, Chester R. Ramsey, Alexander Usynin 👨🔬
Affiliation: Thompson Proton Center, National Institutes for Quantum Science and Technology, Thompson Cancer Survival Center 🌍
Purpose: Four-dimensional CT imaging (4DCT) has become a standard tool for managing respiratory motion in radiation therapy. However, many treatment delivery systems and most diagnostic CT scanners lack the ability to acquire and reconstruct 4D phase images. The purpose of this study was to evaluate an AI-based machine learning technique that generates synthetic 4DCT from a single 3DCT dataset. Synthetic 4DCT could allow the estimation of organ motion before treatment using 3DCT studies and during treatment using IGRT imaging.
Methods: A deep neural network (DNN) was trained using thoracic 3DCT and 4DCT images. The DNN generates synthetic 4DCT images by predicting deformation vector fields (DVFs) from a single input 3DCT dataset. Testing was performed using 4DCT images obtained from an institution independent of the training datasets. The DNN utilized a 3D convolutional autoencoder with shortcut connections to calculate DVFs between exhale (T50) and other respiratory phases (Tn) through B-spline-based deformable image registration. The resulting synthetic 4DCT images were evaluated by comparing individual phase images, GTV volumes derived from the MIP, and HU differences on the average intensity projection against actual 4DCT images.
Results: Synthetic 10-phase 4DCT images were generated for five patients with six target volumes. Actual GTV volumes from the MIP ranged from 3.4 to 406.9 cc, while synthetic MIP GTV volumes ranged from 3.3 to 405.7 cc, with differences within 1.5 cc or 10%. Mean HU values for the AIP were within 11.2 for the external contour, 10.9 for the heart, 104.3 for the liver, and 42.6 for the lungs. Superior-inferior motion was slightly greater in synthetic 4DCT images compared to actual 4DCT images.
Conclusion: Synthetic 4DCT images demonstrated excellent agreement with actual 4DCT images in terms of GTV volumes, HU values, and motion characteristics. The overall accuracy suggests that this approach can effectively simulate respiratory motion.