Flexible Synthetic CT Generation Using Nnsyn Deep Learning Network with Single- and Multi-Sequence Clinical MRI
Abstract
Purpose
Magnetic resonance (MR)–only radiotherapy planning requires accurate synthetic CT (sCT) generation from images acquired using standard clinical MRI simulation protocols. However, MRI acquisition protocols vary substantially across anatomical sites, and many existing sCT models reply on a designated sequence (e.g., T1-weighted mDixon) necessitating inclusion of this sequence in all site-specific protocols. This requirement is suboptimal, as it extends simulation scan time by about 5 minutes per patient. In this study, we developed and evaluated a site-specific sCT generation method that uses only routinely acquired clinical MR sequences and supports heterogeneous single- and multi-sequence inputs across multiple anatomical sites.
Methods
A deep learning–based sCT generation model adapted from the no-new synthetic network (nnsyn) architecture was developed with a flexible input design to accommodate heterogeneous MR sequences. Separate models were trained for each anatomical site using a unified network structure and training strategy. Datasets included prostate (90 patients), abdomen (83 patients), and brain (62 patients). Prostate cases utilized a single standard MR sequence, while abdominal and brain cases used two routinely acquired MR sequences. sCT quality was quantitatively evaluated for each site using voxel-wise mean absolute error (MAE) and peak signal-to-noise ratio (PSNR), computed within the body contour and bony regions.
Results
For abdominal and prostate sites, the proposed model consistently outperformed an in-house cycle generative adversarial network (CycleGAN) based baseline, achieving lower MAE in both soft-tissue and bony regions. Qualitative evaluation demonstrated improved preservation of bony anatomy. For brain cases, the proposed model achieved stable image quality metrics and preserved fine anatomical details, including bone and sinonasal structures.
Conclusion
This study demonstrates the feasibility of flexible, site-specific sCT generation using only standard MR acquisitions with both single- and multi-sequence inputs. Although dosimetric validation is ongoing, prior-model dosimetric benchmarking supports this approach for MR-only radiotherapy planning across multiple anatomical sites.