Author: So Hyun Ahn, Chris Beltran, Byongsu Choi, Jeong Heon Kim, Jin Sung Kim, Bo Lu, Justin Chunjoo Park, Bongyong Song, Jun Tan 👨🔬
Affiliation: Mayo Clinic, Department of Radiation Oncology, Yonsei Cancer Center, Heavy Ion Therapy Research Institute, Yonsei University College of Medicine, Ewha Medical Research Institute, Ewha Womans University College of Medicine, UC San Diego, Yonsei University College of Medicine 🌍
Purpose:
Cone-beam computed tomography (CBCT) is widely used in IGRT for patient positioning but suffers from low resolution and poor soft tissue contrast. Synthetic CT (sCT) generated from CBCT addresses these issues, improving dose calculations and treatment planning. This study proposes High-Quality PatchNet (HQ-PatchNet), which uses high-SSIM CBCT patches and combines patch-based and whole-image learning to enhance resolution, preserve details, and improve sCT accuracy.
Methods:
This study utilized 70 paired CBCT and CT datasets to evaluate three deep learning architectures using patch-based training approaches, with high-quality patches selected to enhance performance:
3D U-Net
3D GAN (Generative Adversarial Network)
3D CycleGAN
For patch-based approaches, 3D patches (16x16x16 voxels) with SSIM > 0.8 were extracted, aligned with CT images using rigid registration, and used for training, while low-SSIM patches were reserved for testing and validation.
HQ-PatchNet synthesized synthetic CT images by combining high-quality patch outputs, with trilinear interpolation applied during post-processing to reconstruct complete volumes. Performance was evaluated using MAE, SSIM, PSNR, and NCC metrics.
Results:
The evaluation metrics for the original CBCT and the top-performing model (HQ-PatchNet using U-Net with high-quality patches) are as follows:
Original CBCT:
MAE: 45.0218 ± 11.8342
NCC: 0.6663 ± 0.0696
SSIM: 0.7324 ± 0.0748
PSNR: 29.6992 ± 3.3819 dB
HQ-PatchNet (3D U-Net):
MAE: 27.6267 ± 9.0461
NCC: 0.9451 ± 0.0238
SSIM: 0.9259 ± 0.0508
PSNR: 37.3627 ± 5.9919 dB
The HQ-PatchNet model significantly outperformed the original CBCT across all key metrics, demonstrating superior accuracy, structural similarity, and image quality. Detailed qualitative and quantitative results for other models are available in the supporting documents.
Conclusion:
The HQ-PatchNet model combines high-SSIM patches with U-Net architecture to generate high-quality synthetic CT images from CBCT data. It outperformed patch-based implementations of GAN and CycleGAN, highlighting its potential to improve adaptive radiotherapy, particularly for head and neck cancer treatment planning.