% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@ARTICLE{Xiao:305638,
author = {F. Xiao and D. Radonic and N. Wahl$^*$ and N. Delopoulos
and A. Thummerer and S. Corradini and C. Belka$^*$ and G.
Dedes and C. Kurz and G. Landry},
title = {{D}eep learning-based synthetic-{CT}-free photon dose
calculation in {MR}-guided radiotherapy: {A}
proof-of-concept study.},
journal = {Medical physics},
volume = {52},
number = {11},
issn = {0094-2405},
address = {Hoboken, NJ},
publisher = {Wiley},
reportid = {DKFZ-2025-02280},
pages = {e70106},
year = {2025},
abstract = {In magnetic resonance imaging (MRI)-guided online adaptive
radiotherapy, MRI lacks tissue attenuation information
necessary for accurate dose calculations. Although deep
learning (DL)-based synthetic computed tomography (CT)
generation models have been developed to obtain CT density
information from MRI, they usually do not meet the
requirement of real-time plan adaptation.We propose a
DL-based photon dose calculation method directly on 0.35 T
MRI to skip synthetic CT generation and show its feasibility
for prostate patient cases.The 0.35 T planning MRI and
deformed planning CT (registered to the planning MRI) of 34
prostate cancer patients treated with a 0.35 T magnetic
resonance-linear accelerator (MR-Linac) were collected. The
air cavities (ACs) in the abdominopelvic area of the
deformed CT images were corrected based on manual AC
contouring on the MRI. Monte Carlo (MC) dose simulations
under a 0.35 T magnetic field were performed on the
corrected CT images. All photon beams were simulated using a
uniform field size of 1 cm × 1 cm $1\,\text{cm} \times
1\,\text{cm}$ . 10 800 beams were simulated with 5 × 10 6
$5\times 10^6$ initial photons for training (20 patients)
and 2160 beams with 5 × 10 7 $5\times 10^7$ photons for
validation (4 patients). For testing, 1080 beams shooting
through the planning target volume (PTV) in 10 patients and
five optimized nine-field intensity-modulated plans were
simulated with 5 × 10 7 $5\times 10^7$ photons. 3D MRI
cuboids covering the photon beams were input into a Unet
model to predict AC segmentation, and 3D MRI and predicted
AC cuboids were input into a long short-term memory (LSTM)
model for beam's eye view (BEV) processing to predict dose.
The gamma passing rate γ pr $\gamma _{\mathrm{pr}}$
(2\%/2mm, D > 10 \% D max $D>10\\%D_{\mathrm{max}}$ ), beam
dose profiles of single beams and dose volume histogram
(DVH) of intensity-modulated plans were evaluated.The test
results for all photon beams from the proposed models
demonstrated a mean γ pr $\gamma _{\mathrm{pr}}$ above
99.50\%. The five treatment plans recalculated by the DL
model each achieved γ pr $\gamma _{\mathrm{pr}}$ values
exceeding 99.80\%. Additionally, the model's inference time
was approximately 12 ms per photon beam.The proposed method
showed that DL-based dose calculation directly on MRI is
feasible for prostate cases, which has the potential to
simplify the procedure for MRI-only workflows and can be
beneficial for real-time plan adaptation.},
keywords = {Deep Learning / Radiotherapy, Image-Guided: methods /
Humans / Magnetic Resonance Imaging / Photons: therapeutic
use / Radiotherapy Planning, Computer-Assisted: methods /
Proof of Concept Study / Radiotherapy Dosage / Prostatic
Neoplasms: radiotherapy / Prostatic Neoplasms: diagnostic
imaging / Radiation Dosage / Male / Monte Carlo Method /
Tomography, X-Ray Computed / LSTM (Other) / MRI (Other) /
deep learning (Other) / dose calculation (Other) / online
MR‐guided photon therapy (Other)},
cin = {E040 / MU01},
ddc = {610},
cid = {I:(DE-He78)E040-20160331 / I:(DE-He78)MU01-20160331},
pnm = {315 - Bildgebung und Radioonkologie (POF4-315)},
pid = {G:(DE-HGF)POF4-315},
typ = {PUB:(DE-HGF)16},
pubmed = {pmid:41186921},
pmc = {pmc:PMC12584934},
doi = {10.1002/mp.70106},
url = {https://inrepo02.dkfz.de/record/305638},
}