High-Resolution CT Imaging Using a Few-Shot Foundation Model
Abstract
Purpose
High-resolution CT is important for radiotherapy contouring and dose calculation. However, acquiring high through-plane resolution is often constrained by radiation dose, motion artifacts, and scanner limitations. This study aims to generate high-resolution (HR) CT images from low-resolution (LR) inputs using a few-shot foundation model.
Methods
We propose an image-domain, super-resolution approach based on a DINOv3-derived foundation model (HR-DINOv3) to reconstruct high through-plane resolution CT from LR inputs. HR-DINOv3 was pretrained on CT-3M, a curated dataset containing 3.87M CT slices. For finetuning, over 35,000 sagittal slices from 197 liver patients were used. LR inputs were created by downsampling ground truth images by a factor of three along the superior–inferior axis. Performance was quantified using Hounsfield Unit (HU) mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM). HR-DINOv3 was compared against (1) a standard DINOv3 model, (2) CycleGAN, and (3) clinically used interpolation. Statistical comparisons were performed using two-sample t-tests.
Results
HR-DINOv3 achieved an MAE of 8.84 ± 4.07 HU, whereas DINOv3 achieved 11.42 ± 3.95 HU, CycleGAN achieved 21.43 ± 6.78 HU, and interpolation yielded 8.69 ± 4.26 HU. HR-DINOv3 attained a PSNR of 43.22 ± 4.81 dB, compared to 40.93 ± 3.28 dB for DINOv3, 35.82 ± 2.83 dB for CycleGAN, and 40.15 ± 3.87 dB for interpolation. For SSIM, HR-DINOv3 achieved 98.06 ± 1.87%, compared to 97.35 ± 2.11% for DINOv3, 93.84 ± 2.93% for CycleGAN, and 97.45 ± 2.13% for interpolation. Relative to HR-DINOv3, all performance differences were statistically significant (p < 0.05), except for MAE between HR-DINOv3 and interpolation (p = 0.13).
Conclusion
DINOv3-based foundation models can produce high-fidelity through-plane CT reconstructions with improved structural similarity and reduced HU error relative to common baselines. These gains may support more accurate contouring and dose calculation in radiotherapy treatment planning.