Source-linked AI summary
Generating Synthesized Computed Tomography (CT) from Cone-Beam Computed Tomography (CBCT) using CycleGAN for Adaptive Radiation Therapy
Xiao Liang, Liyuan Chen, Dan Nguyen, Zhiguo Zhou, Xuejun Gu, Ming Yang, Jing Wang, Steve Jiang
TL;DR
CBCT is clinically available for adaptive radiation therapy but suffers from artifacts and inaccurate HU values, while paired CT-CBCT data for supervised training are difficult to acquire. The paper uses unpaired CycleGAN translation to generate sCT images, which improve image quality and anatomical accuracy relative to existing approaches.
Problem
CBCT images are widely available for ART but contain artifacts and inaccurate HU values, while exactly matched CT-CBCT pairs are difficult to acquire for supervised learning.
Method
A CycleGAN learns unsupervised translation from CBCT to CT-like images using unpaired datasets and cycle consistency, aiming to preserve CBCT anatomy while correcting HU values.
Results
CycleGAN achieved the best visual and quantitative image quality among the evaluated models, removing artifacts, correcting HU values, and outperforming DIR in phantom anatomical accuracy.
Takeaways & Limitations
CycleGAN-generated sCT images can support more accurate segmentation and dose calculation in ART while retaining CBCT anatomy.
Takeaways & Limitations
The approach still requires improvement for metal artifacts, exponential edge gradients, and image truncation, and its evaluation used dpCT as the patient-study HU reference.
Abstract
from arXiv · showhide
Cone beam computed tomography (CBCT) images can be used for dose calculation in adaptive radiation therapy (ART). The main challenges are the large artefacts and inaccurate Hounsfield unit (HU) values. Currently, deformed planning CT images are often used for this purpose, although anatomical accuracy might be a concern. Ideally, we would like to convert CBCT images to CT images with artifacts removed or greatly reduced and HU values corrected while keeping the anatomical accuracy. Recently, deep learning has achieved great success in image-to-image translation tasks. It is very difficult to acquire paired CT and CBCT images with exactly matching anatomy for supervised training. To overcome this limitation, we developed and tested a cycle generative adversarial network (CycleGAN) which is an unsupervised learning method and does not require paired training datasets to synthesize CT images from CBCT images. The synthesized CT (sCT) images have been compared with the deformed planning CT (dpCT) showing visual and quantitative similarity with artifacts being removed and HU value errors being reduced from 71.78 HU to 27.98 HU. Dose calculation accuracy using sCT images has been improved over the original CBCT images, with the average Gamma Index passing rate increased from 95.4% to 97.4% for 1 mm/1% criteria. A deformable phantom study has been conducted and demonstrated better anatomical accuracy for sCT over dpCT.
1. Introduction
Adaptive radiation therapy can use CBCT to track treatment-course anatomy, but CBCT-to-CT conversion is limited by artifacts, inaccurate HU values, and difficulty acquiring paired images with matching anatomy. The paper therefore applies unpaired CycleGAN translation to synthesize CT-like images while retaining CBCT anatomy.
- Clinical motivation: Anatomical changes during radiotherapy can compromise tumor coverage and increase dose to organs at risk.Reported changes include tumor-volume loss, mass displacement, and increased spinal-cord and brainstem doses.
- Clinical motivation: ART studies show dosimetric benefits from tracking anatomy with daily CT or CBCT and replanning when necessary.CBCT is more commonly available in clinics than CT-on-rails.
- Prior approaches: Existing CBCT scatter-correction approaches—including analytical, Monte Carlo, CT-prior, histogram-matching, and learning-based methods—have achieved limited success and often require paired datasets.Paired CBCT and CT images must share exactly matching anatomy, which is difficult to acquire.
- Deep-learning background: Deep learning supports image-to-image translation through supervised paired training or unsupervised training with unpaired images.GANs and CycleGAN are identified as unsupervised translation approaches.
- Study aim: The paper aims to synthesize CT-quality images from CBCT while preserving CBCT anatomy despite the difficulty of obtaining paired training data.CycleGAN is applied to learn translation between CBCT and CT domains without paired images.
2. Materials and Methods
The study uses an unpaired CycleGAN to translate CBCT images into CT-like images while preserving anatomy, combining adversarial, cycle-consistency, and identity losses. It evaluates HU and anatomical accuracy against deformed planning CT and a deformable phantom.
- Data: 13 patients supplied training and validation data, while 4 patients supplied testing data; CT images were resampled and aligned to CBCT resolution.CBCT and CT images were cropped to 512 × 512 axial slices after preprocessing.
- CycleGAN architecture: CycleGAN uses two generators and two discriminators to translate between CBCT and CT domains.GeneratorA maps CBCT to CT, GeneratorB maps CT to CBCT, and the discriminators distinguish real from generated images.
- Generators and discriminators: U-Net generators preserve 512 × 512 × 1 input and output dimensions, while 142 × 142 patchGAN discriminators produce 32 × 32 × 1 outputs.The 142 × 142 receptive field was selected for 512 × 512 image resolution.
- Loss function: Adversarial, cycle-consistency, and identity losses respectively match target-domain distributions, constrain inverse mappings, and preserve CT and CBCT gray levels.Cycle consistency minimizes differences between original and reconstructed CT or CBCT images, while identity mapping preserves HU values.
- Evaluation: The study tests whether sCT improves HU accuracy over CBCT and anatomical accuracy over dpCT using patient data and a deformable H&N phantom.For patient data, dpCT serves as the HU reference; the phantom compares sCT and dpCT anatomy against corresponding CT images.
3. Results
CycleGAN-generated sCT images reduced CBCT artifacts and improved HU agreement with dpCT, while producing dose distributions closer to dpCT. In a deformable phantom, sCT also preserved anatomy more accurately than dpCT.
- Scatter removal and HU improvement: sCT images greatly reduced CBCT artifacts while keeping the same anatomy, whereas dpCT showed slightly different anatomy distributions.
- Scatter removal and HU improvement: The sCT-dpCT regression line closely followed the 45-degree reference line, indicating improved HU value accuracy over CBCT.The reported slopes were 0.98 for sCT and 0.88 for CBCT.
- Scatter removal and HU improvement: 27.98 HU and 86.71 HU were the sCT MAE and RMSE, decreased from 71.78 HU and 167.22 HU for CBCT.SSIM and PSNR for sCT increased to 0.85 and 30.67 from 0.77 and 25.22 for CBCT against CT.
- Scatter removal and HU improvement: sCT HU profiles were corrected to CT HU values in both bone-and-soft-tissue and soft-tissue-only regions, while CBCT profiles were noisy in soft tissue.
- Dose distribution comparison: 97.40% was the mean 3D gamma pass rate for sCT dose distributions versus 95.42% for CBCT under 1 mm/1% criteria.The sCT dose distributions were visually closer to dpCT, while CBCT showed obvious discrepancies.
- Phantom study: In the deformable phantom, sCT had lower MAE and RMSE and higher SSIM and PSNR against CT2 than dpCT.sCT preserved CT2 anatomy while removing CBCT2 artifacts, whereas DIR-based dpCT did not capture the anatomical change as well.
4. Discussion and Conclusions
The CycleGAN model generates sCT images from unpaired CBCT and CT data, removing artifacts and correcting HU values while preserving CBCT anatomy. Across patient and phantom evaluations, sCT improved image quality and anatomical accuracy relative to relevant baselines, while remaining subject to artifact and truncation limitations.
- Unpaired CBCT and CT images enable CycleGAN training without requiring datasets with exactly matching anatomy, which are difficult to acquire.
- CycleGAN removes most CBCT artifacts and corrects HU values, using dpCT images as the reference for HU evaluation.
- MAE decreased from 6.98 HU with DIR to 4.66 HU with CycleGAN, while SSIM increased from 0.91 to 0.95 in the deformable phantom study.
- Compared with DCGAN and progressive growing GANs, CycleGAN achieved the best visual and quantitative image quality.DCGAN removed fewer artifacts, while progressive growing GANs generated false anatomy in truncated regions.
- sCT improved HU accuracy over CBCT while preserving CBCT anatomy and outperforming DIR in anatomical accuracy, supporting more accurate segmentation and dose calculation in ART.
- Future work must improve performance on metal artifacts and exponential edge gradients, address truncation from limited field of view, and test other anatomical sites.