Synthetic CT Generation of the Pelvis in Patients With Cervical Cancer: A Single Input Approach Using Generative Adversarial Network

Atallah Baydoun, Ke Xu, Jin Uk Heo, Huan Yang, Feifei Zhou, Latoya A. Bethell, Elisha T. Fredman, Rodney J. Ellis, Tarun Kanti Podder, Melanie S. Traughber, Raj M. Paspulati, Pengjiang Qian, Bryan J. Traughber, Raymond F. Muzic Jr.IEEE Access 2021 · DOI

TL;DR

This paper introduces a novel deep learning framework, sU-cGAN, for generating synthetic CT (sCT) images from a single T2-weighted MRI sequence in patients with cervical cancer. This approach aims to streamline radiation therapy planning and PET/MR attenuation correction by eliminating the need for a separate CT scan, thereby reducing costs and patient radiation exposure.

What Problem Does This Paper Address?

Precision medicine in oncology, particularly for cervical cancer, relies on multi-modality imaging (PET, MR, CT) for accurate diagnosis and treatment planning. While MR offers excellent soft-tissue contrast and avoids radiation, it lacks the electron density information crucial for radiation therapy dose calculations and PET attenuation correction. Traditional workflows require both MR and CT scans, leading to increased costs, patient discomfort, and potential registration errors between the modalities. Existing methods for generating synthetic CT (sCT) from MR often require multiple MR sequences or have limitations in accuracy. This paper addresses the need for a more efficient and accurate sCT generation method that utilizes a single, commonly acquired MR sequence.

Methods

The authors propose a novel framework called sU-cGAN (shallow U-Net conditional Generative Adversarial Network). The core innovation lies in using a shallow U-Net (sU-Net) with an encoder-decoder depth of 2 as the generator within a conditional GAN (cGAN) architecture. This simplified generator architecture reduces the number of trainable parameters, mitigating the vanishing gradient problem and requiring less training data. The input to the sU-cGAN is a single-channel, T2-weighted TSE-SSH MR sequence, which is routinely acquired for radiological diagnosis and staging in cervical cancer. The discriminator is a standard convolutional network. The loss function combines the cGAN loss with an L1 loss to encourage pixel-wise similarity between the generated sCT and the ground truth CT.

Datasets

The study utilized a dataset of 11 adult female patients with biopsy-proven cervical cancer who underwent both planning CT and PET/MR scans between June 2015 and June 2018 at University Hospitals Cleveland Medical Center. The MR images were T2-weighted, Turbo Spin Echo Single Shot (TSE-SSH) sequences, and CT images were acquired using a Philips Brilliance 16 multislice CT scanner. The authors do not explicitly state the creation of a new benchmark dataset; rather, they use their collected data for training and evaluation.

Key Insights and Ideas

  • Simplification is Powerful: The use of a shallow U-Net (sU-Net) with a depth of 2, instead of deeper architectures, effectively reduces model complexity and computational burden without sacrificing accuracy for this specific task. This challenges the "deeper is always better" paradigm in deep learning.
  • Single Input Efficacy: Demonstrating that a single, standard MR sequence (T2-weighted TSE-SSH) is sufficient for generating accurate sCT for pelvic imaging is a significant step towards a more streamlined workflow.
  • Clinical Workflow Integration: The proposed method directly leverages existing clinical imaging protocols, making it more readily adoptable in clinical practice by avoiding extra acquisition time or specialized sequences.

Key Results

The sU-cGAN framework achieved superior performance compared to other tested methods (U-cGAN, VGG16, and ResNet) in generating synthetic CT images for the female pelvis.

  • Quantitative Metrics: sU-cGAN achieved the lowest mean Absolute Prediction Error (MAPE) of 72.25 ± 25.42 HU and the lowest mean Root Mean Square Deviation (RMSE) of 115.74 ± 21.84 HU. It also achieved the highest Peak Signal-to-Noise Ratio (PSNR) of 63.41 ± 1.67 and the highest Structural Similarity Index Measure (SSIM) of 0.839 ± 0.044.
  • Qualitative Assessment: Visual inspection showed that sU-cGAN and U-cGAN maintained overall anatomical topography, outperforming the blurred images from VGG16 and ResNet. sU-cGAN was visually more similar to the measured CT than U-cGAN.
  • Efficiency: sU-cGAN demonstrated significantly faster training times (7-8 hours) compared to U-cGAN (11 hours), and extremely fast prediction times (around 1.89 seconds per slice).

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Dataset Size: The study was conducted on a relatively small dataset of 11 patients.
  • Pelvic Area Focus: The current work is limited to the pelvic region.
  • PET SUV Validation: The validation of PET standardized uptake value (SUV) is not included.

Future work suggested includes:

  • Generalizability: Extending the framework to other anatomical regions like the brain, head and neck, thorax, and abdomen.
  • 3D Architecture: Developing a 3D sU-Net to leverage spatial relationships for potentially improved results.
  • Larger Datasets: Acquiring more data for both retrospective and prospective studies to further optimize the model.
  • Clinical Validation: Further studies are needed to assess the impact of sCT quality on PET attenuation correction and radiation therapy dose planning, although initial findings suggest it is comparable to existing methods.

Summary updated

BibTeX
@article{baydoun2021synthetic,
  author = {Atallah Baydoun and Ke Xu and Jin Uk Heo and Huan Yang and Feifei Zhou and Latoya A. Bethell and Elisha T. Fredman and Rodney J. Ellis and Tarun Kanti Podder and Melanie S. Traughber and Raj M. Paspulati and Pengjiang Qian and Bryan J. Traughber and Raymond F. Muzic Jr.},
  title = {{Synthetic CT Generation of the Pelvis in Patients With Cervical Cancer: A Single Input Approach Using Generative Adversarial Network}},
  journal = {IEEE Access},
  year = {2021},
  doi = {10.1109/access.2021.3049781},
}
Figure