Related Experiment Video
Updated: May 15, 2026
08:17
Dynamic Lung Tumor Tracking for Stereotactic Ablative Body Radiation Therapy
Published on: June 7, 2015
Deep learning-based intraluminal gas modeling for anatomically accurate synthetic CT in MRI-based radiation therapy
Braian Adair Maldonado Luna1, Gerardo Uriel Perez Rojas1, René Eduardo Rodríguez Pérez1
1Faculty of Physical and Mathematical Sciences, Benemérita Universidad Autónoma de Puebla, Avenida San Claudio y 18 Sur, Puebla, Puebla 72570, Mexico.
Biomedical Physics & Engineering Express
|May 13, 2026
Summary
A new two-stage deep learning framework significantly improves the accuracy of intraluminal gas in synthetic CT (sCT) images, overcoming a key challenge in MRI-only radiotherapy simulation.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiotherapy Physics
Background:
- Stochastic bowel and rectal gas in MRI/CT scans complicates deformable image registration (DIR) for synthetic CT (sCT) generation.
- Manual corrections are often required, hindering efficient MRI-only simulation workflows.
Purpose of the Study:
- To propose and evaluate a deep learning framework that eliminates the need for DIR in sCT generation.
- To improve the definition of intraluminal gas within sCT images.
- To enhance the dosimetric reliability of MRI-only radiotherapy planning.
Main Methods:
- A two-stage Generative Adversarial Network (GAN) framework was developed, utilizing CycleGAN or pix2pix for initial MRI-to-segmented map image conversion, followed by pix2pix for sCT generation.
- Ground-truth gas cavities were identified using convolutional neural network (CNN)-based autosegmentation and manual verification.
- The framework was trained on 60 patients and compared against a single-stage MRI-to-sCT model using Dice-Sørensen coefficient and Mean Absolute Error (MAE).
Main Results:
- The two-stage framework achieved a significantly higher gas cavity Dice-Sørensen coefficient (0.67 ± 0.07 with CycleGAN) compared to single-stage methods (0.57 ± 0.11).
- While Mean Absolute Error (MAE) showed minor variations, dosimetric analysis revealed excellent agreement for the planning target volume (PTV D95%) with minimal differences between one-stage and two-stage models.
- Both models demonstrated negligible deviations in organ-at-risk (OAR) volumetric metrics.
Conclusions:
- The proposed two-stage framework substantially improves the accuracy of intraluminal gas representation in sCT, nearly tripling cavity definition accuracy over single-stage approaches.
- This method effectively addresses a significant bottleneck in MRI-only simulation, maintaining tissue fidelity.
- Further prospective studies are warranted to evaluate the clinical impact on adaptive radiotherapy.
