3030 - Diffusion-Based MR-Equivalent Image Generation from CBCT for Brain Tumor Radiotherapy

Purpose/Objective(s):

The purpose of this study was to generate MR-equivalent images from cone-beam CT (CBCT) using a conditional denoising diffusion probabilistic model (DDPM) for brain tumor radiotherapy and to evaluate whether this approach improves soft-tissue representation compared with conventional CBCT.

Materials/Methods:

A total of 160 patients with brain tumors treated at a single institution were retrospectively included. Planning CT (pCT) and contrast-enhanced T1-weighted MR (CE-T1wMR) images were acquired for each patient. Digitally reconstructed radiographs were generated from pCT, and synthetic CBCT (sCBCT) images were then reconstructed using the Feldkamp–Davis–Kress algorithm. CE-T1wMR, sCBCT, and pCT images were rigidly registered. Window width/level settings of 120/40 HU and 2000/0 HU were applied to pCT and sCBCT images, respectively. Intensity values were normalized to the range of -1 to 1, and all images were resampled to a unified matrix size of 128 × 128 × 256. To suppress background regions and mitigate MR-specific intensity inhomogeneity, an intracranial contour-based mask was applied to all modalities. The dataset was randomly divided into training (n=135), validation (n=5), and test (n=20) cohorts. Model training was performed using axial 2D slices extracted from paired 3D volumes, limited to slices containing the gross tumor volume. MR synthesis was conducted under two input conditions: sCBCT alone (Scenario 1) and combined sCBCT+pCT (Scenario 2), using an L1 loss function. Generated MR images were quantitatively evaluated against reference MR using the structural similarity index measure (SSIM) and normalized cross-correlation (NCC).

Results:

Median SSIM increased from 0.51 for sCBCT to 0.72 in Scenario 1 and 0.75 in Scenario 2. Likewise, median NCC increased from 0.70 for sCBCT to 0.84 and 0.85, respectively. All differences were statistically significant, with Scenario 2 achieving significantly higher SSIM and NCC than Scenario 1 (p < 0.05).

Conclusion:

The proposed conditional DDPM generated MR-equivalent images from pCT and sCBCT with significantly improved structural similarity and correlation relative to reference MR images. These results indicate that DDPM-based CBCT-to-MR synthesis has the potential to enhance soft-tissue visualization and facilitate more accurate image registration in brain tumor radiotherapy.