aDiner: Adaptive Dynamic Implicit Neural Representation for Dynamic CBCT Imaging.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Cone-beam Computed Tomography (CBCT) is crucial for lung cancer radiotherapy, but dynamic reconstruction for motion is challenging.
• Existing supervised methods require paired data and reconstruct images at discrete phases, limiting temporal resolution.
• Dynamic reconstruction from single projections is an ill-posed inverse problem.

Purpose of the Study:

• To develop a self-supervised learning method for high-fidelity dynamic CBCT reconstruction at each timepoint.
• To improve the accuracy and robustness of lung imaging in image-guided radiotherapy.
• To address limitations of existing supervised methods in dynamic CBCT reconstruction.

Main Methods:

• Proposed adaptive Dynamic implicit neural representation (aDiner), a self-supervised approach.
• Introduced adaptive composite representation (ACR) with inter- and intra-cycle encoding (iICE) for temporal coherence.
• Developed probability-based feature fusion (PFF) and motion-guided coarse-to-fine sampling (MCS) including adaptive projection-ray sampling (APS) and adaptive deformed-point sampling (ADS).

Main Results:

• aDiner achieved accurate and robust reconstruction of dynamic CBCTs for each timepoint across simulated, phantom, and clinical datasets.
• The method effectively captured lung motion during both regular and irregular respiratory cycles.
• Demonstrated superior reconstruction quality by separately estimating static and dynamic regions and leveraging motion similarities.

Conclusions:

• aDiner offers a novel self-supervised solution for dynamic CBCT reconstruction in lung radiotherapy.
• The proposed ACR and MCS strategies enhance temporal coherence and motion estimation accuracy.
• This method holds potential for improving image-guided radiotherapy for lung cancer patients by providing continuous motion information.