Author: Matthew Brown, Yushi Chang, Jinhyuk Choi, William Silva Mendes, Lei Ren, Aman Sangal, William Paul Segars, Phuoc Tran, Hualiang Zhong 👨🔬
Affiliation: University of Maryland School of Medicine, Department of Radiation Oncology, University of Maryland School of Medicine, Carl E. Ravin Advanced Imaging Laboratories and Center for Virtual Imaging Trials, Duke University Medical Center, Department of Radiation Oncology, Medical College of Wisconsin 🌍
Purpose: Digital phantoms like XCAT are essential for imaging and treatment optimization in radiology and radiation oncology. However, the lack of realistic textures (HU distribution) in XCAT limits its clinical utility. This study introduces a deep learning-based conditional denoising diffusion probabilistic model (c-DDPM) to generate realistic CT textures, creating a realistic textured XCAT (RT-XCAT) phantom.
Methods: The RT-XCAT phantom was developed using a ControlNet-based architecture and a customized UNet. c-DDPM is trained by first iteratively adding gaussian noise to the original data and then using the neural networks to learn the reverse denoising process. In our study, organ maps were generated from real CT images with piece-wise uniform organ intensity to imitate XCAT. Then the model was trained to synthesize realistic CT images from organ maps with the actual patient CT as the ground-truth. 240 liver cancer patients’ CT data were used for model training. Once trained, the model was applied to the XCAT phantom to generate RT-XCAT with realistic textures. Model performance was evaluated using Structural Similarity Index Metric (SSIM), Peak Signal-to-Noise Ratio (PSNR), and Hounsfield Unit (HU) distribution comparisons between real and simulated CT images. An observer study was performed by two oncologists to further assess RT-XCAT’s clinical realism.
Results: Comparison of simulated and real CT achieved a mean SSIM of 0.88±0.04 and a mean PSNR of 26.30±2.81, indicating high similarity to real CT. In the observer study, 83% of RT-XCAT images were rated as realistic CT compared to less than 1% for XCAT images, marking a significant improvement.
Conclusion: This study demonstrates the first successful application of diffusion models to create a realistic digital phantom in the abdominal region. The RT-XCAT phantom has significant potential for enhancing radiotherapy applications, such as motion management and robust treatment planning, improving the precision of both photon and proton therapies.