Author: Xiaoxue Qian, Hua-Chieh Shao, You Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, UT Southwestern Medical Center 🌍
Purpose:
Limited-angle CBCT (LA-CBCT) reduces imaging time and dose but suffers from under-sampling artifacts. 2D–3D deformable registration addresses this problem by estimating LA-CBCTs from deforming a prior, fully-sampled CT/CBCT using deformation-vector-fields (DVFs) optimized by limited-angle 2D projections. Population-trained neural networks achieve fast inference but face accuracy challenges, especially under changing limited-angle scan directions. We developed a hybrid population-based and patient-specific 2D-3D registration framework (HB-2D3DReg) by leveraging population-based models’ fast inference and patient-specific models’ test-time adaptability to improve the efficiency/accuracy of 2D–3D registration-driven LA-CBCT estimation.
Methods:
HB-2D3DReg synergizes the advantages of population-based and patient-specific approaches while addressing their limitations, with a two-stage approach. First, a population-based 2D-3D registration network was trained via a population dataset in an unsupervised fashion, with a similarity loss defined between digitally-reconstructed-radiographs (DRRs) of the estimated LA-CBCTs and limited-angle 2D projections. Then, a 2D-3D registration network, based on implicit-neural-representation (INR), refined the DVFs solved by the population-based model during test time for each independent testing case. The population-based method accelerates the optimization of the patient-specific INR network, while the patient-specific INR network, in turn, enhances the accuracy of the population-based model.
HB-2D3DReg was evaluated using a 4D-CT lung dataset of 36 cases, 26 of which were used to train the population-based model and 10 for testing. Different limited-angle scan scenarios, featuring varying scan directions and angles, were evaluated.
Results:
HB-2D3DReg attained superior LA-CBCT estimation and registration accuracy. Under an orthogonal-view 90° scan (45° each) with varying scan directions, HB-2D3DReg achieved mean(±S.D.) relative-error of 8.54±2.36% and target-registration-error of 4.19±2.51mm, compared to 15.40±2.41% and 8.52±3.31mm (no-registration), 9.82±2.25% and 6.38±3.05mm (population-model-only), and 9.71±2.33% and 5.01±2.77mm (INR-registration-only). HB-2D3DReg took ~3min at test time, compared to 13min for the INR-only method.
Conclusion:
HB-2D3DReg achieves accurate and robust 2D-3D deformation registration for LA-CBCT estimation, enabling efficient anatomy monitoring to guide radiotherapy.