Author: Hacene Azizi, Abderaouf Behouch, Nabil Maalej, Aamir Raja, Mohamed Lamine Seghier 👨🔬
Affiliation: Khalifa University Biomedical Engineering & Biotechnology, Laboratory of dosing, analysis and characterization in high resolution, Ferhat Abbas Setif 1 University, Khalifa University Physics Department, Khalifa University 🌍
Purpose:
To ensure accurate alignment of cone-beam computed tomography (CBCT) and computed tomography (CT) images for image-guided radiotherapy, this study evaluates a foundation model guidance framework using zero-shot and few-shot strategies for CBCT-CT registration, emphasizing the importance of robust initial rigid or affine registration.
Methods:
The framework uses an adapter layer integrated with the foundation model to represent CBCT and CT volumes in three channels. For sagittal, axial, and coronal views, three statistical measures—standard deviation, maximum, and average—are computed and combined into composite slices. These slices are processed using DINO, a self-supervised Vision Transformer (ViT) model, to extract domain-independent features. Affine transformations are estimated using the matched features for each plane, and a final 3D transformation is derived.
Results:
The proposed framework demonstrates comparable performance with traditional methods in achieving robust domain-invariant CBCT-CT registration.
Conclusion:
The study highlights the significance of robust initial alignment in CBCT-CT registration and demonstrates the potential of foundation model-guided strategies to address global misalignments in image-guided radiotherapy.