Deep learning-based conversion of MRI to synthetic CT for gamma knife radiosurgery – A large-sample study

Purpose Gamma Knife radiosurgery (GKRS) relies on planning computed tomography (CT) for dose calculation, which introduces inherent CT/MR registration errors, unnecessary radiation exposure, and suboptimal soft-tissue contrast. While deep learning-based synthetic CT (sCT) generation from MRI offers a promising solution, existing studies predominantly utilize small sample sizes (typically fewer than 100 cases), which may limit prediction accuracy. This study aims to investigate the impact of a large-sample training dataset on sCT generation accuracy and evaluate its dosimetric performance in frameless GKRS. Methods and Materials: A total of 246 patients treated with GKRS for meningiomas were retrospectively enrolled. Eighteen cases were randomly selected for testing, while the remaining 228 cases were partitioned into a small-sample group (28 cases, M28) and a large-sample group (200 cases, M200) to train a U-Net-based MRI-to-CT translation model. The image synthesis accuracy was evaluated using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) across the whole body, soft tissue, bone, and intra-scalp regions. For the M200 model, dosimetric verification was performed by transferring the sCT and corresponding contours into the treatment planning system. Dose-volume histogram (DVH) parameters for the gross tumor volume (GTV) and various organs at risk (OARs) were calculated and statistically compared with the ground-truth CT plans. Results The M200 model consistently outperformed the M28 model across all evaluated regions and metrics. Specifically, M200 yielded significantly lower MAE (e.g., bone: 126.55 vs. 175.95; body: 18.53 vs. 25.13) and higher SSIM (e.g., bone: 0.91 vs. 0.84) compared to M28. Visual assessment and difference maps confirmed high anatomical fidelity with minimal HU errors, primarily localized at complex bony boundaries. Dosimetric evaluation demonstrated no statistically significant differences in target coverage (D98, D95, D2) or OAR sparing between the original CT and sCT plans (all p > 0.05). The DVH curves for both the GTV and OARs were nearly perfectly superimposed. Conclusions Increasing the training sample size substantially improves the accuracy of MRI-based sCT generation, particularly in bony structures. The sCT generated by the large-sample model provides sufficient HU accuracy and dosimetric reliability for GKRS, offering a viable alternative for MRI-only treatment planning workflows.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11503881/v1
Primary Topic
Advanced Radiotherapy Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Deep learning-based conversion of MRI to synthetic CT for gamma knife radiosurgery – A large-sample study

Mingying Geng, Shengxiu Jiao, Jia Luo, Shuaishuai Xu et al.
Research Square
Advanced Radiotherapy Techniques
preprint

Deep learning-based conversion of MRI to synthetic CT for gamma knife radiosurgery – A large-sample study

Mingying Geng, Shengxiu Jiao, Jia Luo, Shuaishuai Xu, Peng Zhou, Enyu Liu, Wansheng Liu, Jian Li, Jiangfeng Wang
preprint en

Abstract

Purpose Gamma Knife radiosurgery (GKRS) relies on planning computed tomography (CT) for dose calculation, which introduces inherent CT/MR registration errors, unnecessary radiation exposure, and suboptimal soft-tissue contrast. While deep learning-based synthetic CT (sCT) generation from MRI offers a promising solution, existing studies predominantly utilize small sample sizes (typically fewer than 100 cases), which may limit prediction accuracy. This study aims to investigate the impact of a large-sample training dataset on sCT generation accuracy and evaluate its dosimetric performance in frameless GKRS. Methods and Materials: A total of 246 patients treated with GKRS for meningiomas were retrospectively enrolled. Eighteen cases were randomly selected for testing, while the remaining 228 cases were partitioned into a small-sample group (28 cases, M28) and a large-sample group (200 cases, M200) to train a U-Net-based MRI-to-CT translation model. The image synthesis accuracy was evaluated using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM) across the whole body, soft tissue, bone, and intra-scalp regions. For the M200 model, dosimetric verification was performed by transferring the sCT and corresponding contours into the treatment planning system. Dose-volume histogram (DVH) parameters for the gross tumor volume (GTV) and various organs at risk (OARs) were calculated and statistically compared with the ground-truth CT plans. Results The M200 model consistently outperformed the M28 model across all evaluated regions and metrics. Specifically, M200 yielded significantly lower MAE (e.g., bone: 126.55 vs. 175.95; body: 18.53 vs. 25.13) and higher SSIM (e.g., bone: 0.91 vs. 0.84) compared to M28. Visual assessment and difference maps confirmed high anatomical fidelity with minimal HU errors, primarily localized at complex bony boundaries. Dosimetric evaluation demonstrated no statistically significant differences in target coverage (D98, D95, D2) or OAR sparing between the original CT and sCT plans (all p > 0.05). The DVH curves for both the GTV and OARs were nearly perfectly superimposed. Conclusions Increasing the training sample size substantially improves the accuracy of MRI-based sCT generation, particularly in bony structures. The sCT generated by the large-sample model provides sufficient HU accuracy and dosimetric reliability for GKRS, offering a viable alternative for MRI-only treatment planning workflows.

Research Square
Army Medical University (CN), Daping Hospital (CN), Shandong Provincial Hospital (CN), Shandong First Medical University (CN)
Advanced Radiotherapy Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon