A Generalizable Deep Learning Model for Real-Time CT Generation from a Single Projection for Lung Cancer Radiotherapy
Abstract
Purpose
Respiratory during lung cancer radiotherapy causes tumor and organ motion, increasing uncertainties of absorbed doses, while current image-guided methods cannot provide real-time volumetric CT during treatment. We proposed a generalizable deep learning model, SP2V (Single Projection to Volume), for real-time CT generation from a single CBCT projection and evaluated its image quality and anatomical accuracy in comparison with other methods.
Methods
The 4DCT from 150 lung cancer patients were collected. Cone-beam projections at 0° were generated and divided into training (105), validation (15), and testing (30) set. Deformation vector fields (DVFs) were obtained by registering the end-exhale phase to other phases using TransMorph. SP2V, combining CNN and Transformer architectures, takes the reference CT and paired projections as input to estimate DVF for target CT generation. SP2V was compared with 3DUnet and CT-to-CT deformable registration–based methods (Elastix and TransMorph). Loss functions included CT MSE, DVF MSE, and DVF smoothness, with ablation experiments for choosing the optimal training objective. Evaluation metrics included MAE, PSNR, SSIM, Dice and HD95 for GTV, Lung Dice, and the number of voxels with negative Jacobian determinant.
Results
SP2V trained with CT MSE and DVF MSE achieved the best overall performance an inference time of 10 ms, producing smooth deformations (Jacobian < 0 voxels: 24.9), high image quality (MAE: 9.88 HU, PSNR: 37.4 dB, SSIM: 0.971), and accurate anatomy preservation (GTV Dice: 0.813, HD95: 4.56 mm, Lung Dice: 0.975), outperforming 3DUnet across all metrics. Compared with Elastix and TransMorph, SP2V showed better MAE and Lung Dice against Elastix, inferior image quality and GTV accuracy, but more physically consistent DVFs with fewer folding voxels.
Conclusion
SP2V provided real-time generation of high-quality thoracic CT from a single CBCT projection without patient-specific training, with high anatomical accuracy and physically plausible deformations, demonstrating its potential for CBCT-based real-time CT-guided adaptive radiotherapy.