Source-linked AI summary
Dose evaluation of fast synthetic-CT generation using a generative adversarial network for general pelvis MR-only radiotherapy
Matteo Maspero, Mark H. F. Savenije, Anna M. Dinkla, Peter R. Seevinck, Martijn P. W. Intven, Ina M. Jurgenliemk-Schulz, Linda G. W. Kerkmeijer, Cornelis A. T. van den Berg
TL;DR
MR-only radiotherapy needs rapid synthetic-CT generation for dose calculation without relying on CT images. This study evaluated a paired cGAN using multi-contrast MR images and found accurate pelvic dose calculations, with generation completed within seconds.
Problem
MR-only radiotherapy aims to use MR images for patient simulation and dose calculation, motivating rapid synthetic-CT generation for MR-guided workflows.
Method
A paired cGAN generated synthetic-CT images from registered, normalized, multi-contrast MR images, with air cavities explicitly harmonized between CT and MR datasets.
Results
Less than 0.5% average dose difference was reported for MR-based versus CT-based calculations, while generation took within 6 seconds on a GPU and 21 seconds on a CPU.
Takeaways & Limitations
The cGAN generated accurate synthetic-CT images for pelvic prostate, rectal, and cervical cancer cases without retraining, supporting potential MR-guided radiotherapy use.
Takeaways & Limitations
Dosimetric evaluation was restricted to the pelvic area, and assessment was performed without accounting for magnetic-field effects.
Abstract
from arXiv · showhide
To enable magnetic resonance (MR)-only radiotherapy and facilitate modelling of radiation attenuation in humans, synthetic-CT (sCT) images need to be generated. Considering the application of MR-guided radiotherapy and online adaptive replanning, sCT generation should occur within minutes. This work aims at assessing whether an existing deep learning network can rapidly generate sCT images to be used for accurate MR-based dose calculations in the entire pelvis. A study was conducted on data of 91 patients with prostate, rectal and cervical cancer who underwent external beam radiotherapy acquiring both CT and MRI for patients' simulation. Dixon reconstructed water, fat and in-phase images obtained from a conventional dual gradient-recalled echo sequence were used to generate sCT images. A conditional generative adversarial network (cGAN) was trained in a paired fashion on 2D transverse slices of 32 prostate cancer patients. The trained network was tested on the remaining patients to generate sCT images. For 30 patients in the test set, dose recalculations of the clinical plan were performed on sCT images. Dose distributions were evaluated comparing voxel-based dose differences, gamma and dose-volume histogram (DVH) analysis. The sCT generation required 5.6 s and 21 s for a single patient volume on a GPU and CPU, respectively. On average, sCT images resulted in a higher dose to the target of maximum 0.3%. Results suggest that accurate MR-based dose calculation using sCT images generated with a cGAN trained on prostate cancer patients is feasible for the entire pelvis. The sCT generation was sufficiently fast to be integrated into an MR-guided radiotherapy workflow.
1. Introduction
MR-only radiotherapy uses MR images for simulation and dose calculation, but MR alone does not provide the electron-density information needed for radiation attenuation modelling. This study assesses whether a cGAN can rapidly generate pelvic sCT images suitable for MR-based dose calculation, including cancers beyond prostate.
- Motivation: MR-only radiotherapy performs patient simulation and dose calculation using only MR images.The workflow may reduce registration-related target-definition errors and practical burdens.
- Problem: MR images cannot directly support dose calculation because no demonstrated correlation links MRI nuclear magnetic properties to electron density.sCT images are therefore generated to model radiation attenuation for MR-based dose calculation.
- Research gap: Pelvic sCT methods have mainly been evaluated for prostate cancer, while only three studies investigated other pelvic sites.Reported prostate dose deviations were below 2% relative to CT-based calculations.
- Research gap: sCT generation should take minutes or less to support daily replanning during MR-guided radiotherapy.Earlier deep-learning methods generated full 3D volumes in about one minute, but generation-time attention remained limited.
- Study aim: The study evaluates an existing cGAN trained with paired MRI-CT data to generate sCT images from multi-contrast Dixon MR images.Training used prostate cancer patients, while dosimetric evaluation included prostate, rectal and cervical cancer patients.
2.1. Patient data collection
The study collected paired CT and MRI simulation data from 91 patients with prostate, rectal or cervical cancer undergoing external-beam radiotherapy. Imaging protocols, positioning and treatment characteristics varied across these patient groups.
- Cohort: 91 patients were included: 59 with prostate, 18 with rectal and 14 with cervical cancer.All underwent external-beam radiotherapy and had both CT and MRI simulation images.
- Prostate cohort: The prostate cohort comprised 59 patients with low- to high-risk disease and received five-beam 10 MV IMRT.Prescription doses were 35x2.2 Gy to prostate and macroscopic tumour and 35x2.0 Gy to seminal vesicles.
- Rectal cohort: The rectal cohort comprised 18 patients with intermediate- or high-risk T2-T4 disease treated with neoadjuvant radiotherapy.Treatment included short-course 5x5 Gy and long-course 25x2.0 Gy regimens, with some integrated boosts.
- Cervical cohort: The cervical cohort comprised 14 patients with low- to high-risk T1-T4 disease treated using 10 MV VMAT.Dose schemes included 25x1.8 Gy, 25x2.2 Gy and 25x2.3 Gy, with integrated boosts in some patients.
- Imaging protocol: MRI was acquired within 2.5 hours of CT for prostate and rectal patients, whereas cervical imaging intervals reached one week.Patients followed hydration and bladder or rectal preparation instructions before scanning.
2.2. The network
The study used the pix2pix conditional generative adversarial network to generate synthetic CT images from paired MR inputs. Its architecture combines a U-net generator with a PatchGAN discriminator.
- Architecture: The pix2pix cGAN used a 256x256 U-net generator and a 70x70 PatchGAN discriminator.The implementation was adopted from the PyTorch version provided by Isola and colleagues.
- Training objective: Training optimized the cGAN objective cGAN+λ·L1 with λ=100.The combined adversarial and L1 loss was selected to balance sharpness against image artefacts and blurring.
2.3. sCT generation
The sCT pipeline registered and normalized paired CT and Dixon MR images, enforced consistent air-pocket locations, and generated volumetric sCT by stacking transverse cGAN outputs. The resulting volumes were converted for treatment-planning use and evaluated with a separate air-consistency experiment.
- Image pre-processing: CT images were rigidly registered and resampled to MR, then both modalities were cropped to a common field of view and normalized.CT intensities were clipped to [−1000;1047] HU and all images were converted to 8-bit inputs for pix2pix.
- Image pre-processing: Air cavities were filled and assigned −1000 HU according to their MR locations before cGAN training.This CTair preprocessing was introduced to reduce inconsistent air depiction between MRI and generated sCT.
- Network training: Paired cGAN training used transverse slices from 32 randomly selected prostate cancer patients for 200 epochs.Training used a Tesla P100 GPU, batch size one, horizontal flipping and random paired cropping.
- Image generation: The trained generator produced sCT for unused patients by processing transverse slices and stacking them into 3D volumes.Post-processing with DCMTK created DICOM files usable in a treatment-planning system.
- Image generation: A comparison model trained directly on CTreg generated sCTNoAir for 27 prostate test patients.This experiment assessed the impact of enforcing air-location consistency during training.
2.4. Evaluation
The evaluation compared sCT-based images and dose calculations with registered CT references across pelvic radiotherapy cases. Analyses combined image-error metrics, dose differences, gamma passing rates, and DVH-point comparisons.
- Evaluation setup: Network performance was evaluated by measuring cGAN training and generator-inference time on GPU and CPU frameworks.The CPU framework used a quad-core Intel Xeon 3.4 GHz processor.
- Image evaluation: Image accuracy was assessed using MAE and ME between sCT and CTreg within the intersection of automatically thresholded body contours.The body contours were obtained by thresholding CTreg and sCT at −500 HU.
- Image evaluation: For prostate patients, sCTNoAir was additionally compared with CTreg using MAE and ME.This comparison included 27 prostate patients in the test set.
- Dose evaluation: Thirty test patients—10 each with prostate, rectal, or cervical cancer—underwent clinical-plan dose recalculation on sCT images.Calculations used Monaco with a Monte Carlo photon algorithm on a 3x3x3 mm3 grid.
- Dose evaluation: MRI field-of-view limitations prevented dose recalculation for most cervical patients, so only two patients with comparable inferior coverage were evaluated.The limitation reflected MRI acquisition for primary-tumour delineation while some plans included nodal boosts outside the acquired field of view.
- Dose evaluation: Dose differences were computed as CT−sCT, with gamma analysis performed at 3%,3mm and 2%,2mm across prescription-dose thresholds and the cropped body intersection.The analysis excluded dose build-up using a 15 mm crop and evaluated DVH points for targets and relevant organs at risk.
3. Results
The cGAN generated pelvic sCT volumes rapidly and produced image and dose comparisons that were close to CT-based references. Across tumour types, dose differences and DVH-point differences remained small, supporting accurate sCT-based dose calculation.
- Image comparison: 61±9 HU MAE and 2±8 HU ME were obtained on average across the 59-patient test set when comparing sCT with CTreg.The reported image errors were comparable among patients with different tumour locations.
- Image comparison: Comparable MAE and ME were observed with or without enforced air-location consistency, although ME slightly decreased without enforcement.The authors attribute the limited metric impact to air pockets occupying a small fraction of the body contour.
- Dose comparison: All target and organ-at-risk DVH points on sCT-derived plans were within ±2.5% of corresponding CT-derived points.The comparison included PTV, CTV, and relevant organs at risk.
4. Discussion
The cGAN generated sCT images rapidly and supported accurate pelvic MR-based dose calculations across prostate, rectal, and cervical cancer cases, while clinical deployment remains bounded by field-of-view and magnetic-field considerations.
- Dosimetric accuracy: 0.5% average dose difference between CT- and sCT-based calculations was observed after training the cGAN on prostate cancer patients.The result was reported for pelvic patients, including rectal and cervical cases, although cervical clinical-plan recalculation was limited by field-of-view coverage.
- Scope and limitations: The dosimetric assessment was restricted to the pelvic area, and the method would require appropriate retraining for other anatomical locations.
- Novelty and context: The study provides the first reported dosimetric evaluation of deep-learning-generated sCT images in the pelvic region.The reported dose deviations were described as consistent with prior pelvic and prostate studies.
- MR acquisition: Using multi-contrast images from one conventional MR acquisition may reduce sequence requirements, although its effect on sCT quality remains unclear.The authors suggest that specialised sequences might become unnecessary, while retaining the need for high geometric fidelity.
- Scope and limitations: Dose calculations were assessed without a magnetic field, so accuracy for MR-guided radiotherapy with magnetic fields remains to be investigated.Magnetic fields can affect dose distributions, particularly near air cavities.
5. Conclusion
The study found that a cGAN using multi-contrast MR images from one acquisition enabled accurate pelvic sCT-based dose calculations and generated volumes within seconds.
- Conclusion: A cGAN using multi-contrast MR images from a single acquisition enabled accurate dose calculations in prostate, rectal, and cervical cancer patients without retraining.The network was trained on prostate cancer patients and applied across the pelvic region.
- Conclusion: sCT generation required within 6 seconds on GPU and within 21 seconds on CPU, benefiting time-sensitive MR-guided radiotherapy applications.