Planning CT-Guided Dual-Branch Attention GAN for Longitudinal CBCT Outpainting in Adaptive Radiotherapy

Area of Science:

• Medical Imaging
• Radiotherapy Technology
• Artificial Intelligence in Medicine

Background:

• Limited longitudinal field of view in cone-beam computed tomography (CBCT) poses challenges for image-guided adaptive radiotherapy.
• Accurate anatomical representation is crucial for effective treatment planning and delivery.

Purpose of the Study:

• To develop and evaluate a deep learning-based CBCT outpainting framework.
• To synthesize anatomically consistent extensions for limited-FOV CBCT images.
• To improve image quality and anatomical fidelity for adaptive radiotherapy.

Main Methods:

• A dual-branch encoder-decoder deep learning architecture was employed.
• The model integrated CBCT data with edge maps from planning CT (pCT).
• Attention-guided skip connections and late fusion were used for cross-modality integration.

Main Results:

• The framework achieved minimal absolute volume differences for clinical target volume (0.11 ± 0.07 cc) and planning target volume (0.05 ± 0.04 cc).
• Superior perceptual image quality was demonstrated with LPIPS (0.119 ± 0.029) and FID (54.70).
• Qualitative assessments showed improved soft-tissue continuity and preservation of anatomical landmarks.

Conclusions:

• The proposed deep learning framework effectively restores longitudinal CBCT coverage with high anatomical fidelity.
• This technology supports more reliable deformable registration and adaptive radiotherapy workflows.
• It addresses limitations of field-of-view constraints in clinical CBCT applications.