Related Experiment Video
Updated: Mar 17, 2026
Four-Dimensional CT Analysis Using Sequential 3D-3D Registration
Published on: November 23, 2019
Unsupervised deep learning-based 4DCT deformable image registration and carbon-ion 4D dose calculation
Xiaoyan An1,2,3,4,5, Jian Wang1,2,3,6, Jun Zhang1,4
1Institute of Modern Physics, Chinese Academy of Sciences, Lanzhou, China.
Background:
Accurate four-dimensional dose calculation (4DDC) is essential for carbon-ion lung radiotherapy and relies on deformable image registration (DIR). However, conventional DIR methods are computationally intensive, hindering the implementation of online adaptive workflows.
Purpose:
This study investigates the feasibility and efficacy of unsupervised deep learning-based DIR models, TransMatch and VoxelMorph, in accelerating clinical lung four-dimensional computed tomography (4DCT) registration and facilitating accurate carbon-ion 4D dose calculation.
Methods:
A total of 150 clinical lung 4DCT datasets were utilized (120 for training, 20 for validation, and 10 for testing), with a conventional B-spline method serving as the baseline. Registration accuracy was evaluated using the Mean Absolute Error (MAE), Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance (HD95), and Jacobian determinant (|J|). Carbon-ion 4D dose distributions were accumulated using the generated deformation vector fields (DVFs). Dosimetric impacts on the gross tumor volume (GTV) and organs at risk (OARs) were quantified using Dose-Volume Histogram (DVH) metrics under a gating with 6× rescanning scenario.
Results:
TransMatch and VoxelMorph achieved superior registration accuracy with lower MAE, mean DSC > 0.97, and HD95 < 2.5 mm. In DVF analysis, TransMatch and VoxelMorph showed negligible folding rates (<0.02%) and significantly higher Jacobian standard deviations than the B-spline method, indicating superior capability in capturing fine local deformations. Dosimetrically, differences in GTV and OARs metrics between the deep learning and B-spline methods were less than 2% of the prescription dose, falling within clinically acceptable tolerances. Crucially, TransMatch and VoxelMorph models achieved sub-second registration times (<1 s), whereas the conventional B-spline method required more than 10 min.
Conclusions:
TransMatch and VoxelMorph achieve geometric and dosimetric accuracy comparable to the conventional B-spline method for carbon-ion lung radiotherapy while offering substantially higher computational speed, highlighting their potential for real-time adaptive carbon-ion therapy.
