High-Fidelity Abdominal Synthetic CT Generation for MR-Only Radiotherapy on an MR-Linac Using a 2.5D Latent Consistency Model
Abstract
Purpose
Synthetic CT (sCT) provides the electron density information needed for MR-only treatment planning, eliminating CT-MR registration errors and patient burden reduction. There are hitherto no reliable methods for sCT of the abdomen due to respiratory motion, organ deformation, and intestinal gas variations. Generative diffusion models offer the robust capability required to model these complex anatomical variations where traditional methods fail, while the specific architecture of Latent Consistency Models (LCM) ensures the computational efficiency necessary for practical clinical workflows. This study presents a novel LCM with 2.5D multi-slice architecture to generate high-quality sCT images from MR-Linac acquisitions for abdominal radiotherapy planning.
Methods
An LCM was trained using fifty-six abdominal patients who underwent CT simulation and MRI on a 1.5T Elekta Unity MR-Linac using T1-weighted sequences (TR/TE=8/3.6ms, flip angle=8°, matrix=384×384, FOV=270mm, slice thickness=2.2mm). Planning CT from a Philips Brilliance Big Bore CT-SIM served as ground-truth. CT images were clipped to [-1024, 3071] HU and rescaled to [-1, 1]; MR images scaled to [-1, 1]. CT volumes were deformably registered to MRI. The 2.5D input for the LCM utilized five adjacent MRI slices employing tissue-weighted Mean Squared Error (MSE) loss, Modality Independent Neighbourhood Descriptor (MIND) loss, and Learned Perceptual Image Patch Similarity (LPIPS) loss. The algorithm was trained with images in all orientations. Four test patients were held out for validation using mean absolute error (MAE), root mean square error (RMSE) and Normalized cross-correlation (NCC).
Results
The MAE of the generated CTs was 69±6 HU and an RMSE of 117±13 HU. The NCC was 0.96±0.01. Average inference time was under 1minute.
Conclusion
The proposed LCM with 2.5D multi-slice architecture generates clinically viable sCT images from 1.5T MR-Linac acquisitions for abdominal radiotherapy. These results support feasibility of MR-only workflow implementation for abdominal radiotherapy, potentially eliminating systematic registration errors while streamlining clinical workflow.