Author: Rashmi Bhaskara, Oluwaseyi Oderinde 👨🔬
Affiliation: Purdue University 🌍
Purpose: This study proposes a novel approach to overcoming CBCT image quality limitations by developing an improved synthetic CT (sCT) generation method based on a CycleGAN architecture using Vision Transformer (ViT) with an Adaptive Fourier Neural Operator (AFNO).
Methods: A dataset of 20 prostate cancer patients who received stereotactic body radiation therapy (SBRT) was used, consisting of paired CBCT and planning CT (pCT) images. The preprocessing pipeline involved spatial registration, resampling to uniform voxel sizes, and normalization. The model architecture combines a CycleGAN with bidirectional generator networks, where the UNet generator is augmented with a ViT at the bottleneck, and an AFNO mechanism is applied in the Fourier domain. The AFNO's key innovations include its ability to handle varying resolutions, mesh invariance, and efficient capture of long-range dependencies and periodic patterns.
Results: The proposed model showed significant improvements in preserving fine anatomical details and capturing complex image dependencies. The AFNO mechanism facilitated effective processing of global image information while adapting to inter-patient anatomical variations, resulting in more accurate sCT generation. Evaluation using key metrics—Mean Absolute Error (MAE), Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Normalized Cross-Correlation (NCC)—showed that our model outperforms others. Specifically, our model achieved an MAE of 9.71, a PSNR of 37.08 dB, SSIM of 0.97, and NCC of 0.99, confirming the efficacy of the proposed method.
Conclusion: This method, integrating AFNO within the CycleGAN-UNet framework, effectively addresses inferior CBCT image quality concerns. The model's ability to preserve both global and local anatomical features holds promise for improving tumor targeting and supporting clinical decision-making.