Extending longitudinal field-of-view for cone-beam CT based on a novel surface-aware diffusion model.

Abstract

The limited longitudinal field-of-view (FOV) of cone-beam CT (CBCT) impedes accurate patient setup and adaptive radiotherapy (ART). A surface-aware DDIM-based RePaint diffusion framework was proposed to extend the longitudinal CBCT anatomies with inter-fractional changes, by leveraging prior planning CT (pCT) and updated optical surface image (OSI) as prior knowledge. Three CBCTs on different days and one pCT of 41 patients from two centers were retrospectively selected (Center A: 34 training, 4 testing, Trilogy; Center B: 3 testing, Ethos), yielding 102 training pairs and 21 test pairs. A surface-aware DDIM-RePaint model was developed to restore truncated CBCT volumes of the treatment day conditioned on prior pCT and updated OSI inputs. The masked (restored) volumes were quantitatively evaluated using mean absolute error (MAE, HU), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR). Ablation studies were performed to assess the contribution of OSI. On testing cases, DDIM-RePaint achieved MAE = 25.28 ± 2.45 HU, SSIM = 94.89 ± 0.73, PSNR = 23.05 ± 0.83 for Center A, and MAE = 35.20 ± 4.30 HU, SSIM = 91.27 ± 2.50, PSNR = 21.12 ± 0.89 for Center B, both were significantly better than that of fusing pCT and CBCT as the current clinical approach (all P < 0.001). Based on various metrics and extension proportions, ablation results consistently demonstrated the significant contribution of OSI to the accuracy of longitudinal image restoration, which decreased monotonically with the extending volume. The inference time on 12 samples was reduced from 45 minutes to 2.2 minutes by sampler optimization. The proposed surface-aware DDIM-RePaint framework provided an accurate and efficient approach for longitudinal CBCT FOV extension, potentially facilitating patient setup and ART planning without extra imaging dose and hardware upgrade.