Author: Chia-Ho Hua, Jirapat Likitlersuang, Jinsoo Uh 👨🔬
Affiliation: St. Jude Children's Research Hospital 🌍
Purpose: AI-based fast MRI, which reconstructs images from undersampled k-space data, has not yet been tailored for RT planning. This study aims to evaluate the fast MRI performance of our recently proposed adaptive DDPM in comparison with a leading non-adaptive DDPM, focusing on pediatric brain tumors with diversities in size, location, and resection status.
Methods: As a DDPM is unconditionally trained, the generated images are subject to artifacts, unless the inference is properly conditioned. We proposed a DDPM-based fast MRI method that augments the conditioning by adapting the denoiser network to patient-specific prior images, thereby improving the reconstruction of individual anatomic abnormalities. Meanwhile, conventional non-adaptive DDPM for generic images has been refined to suppress artifacts by modulating the backbone and skip connections in the denoiser U-net (Si, 2024), which has been adopted for fast MRI under the name Texture Coordination (TC)-DiffuRecon (Zhang, 2024). We compared our fast MRI method with TC-DiffuRecon for the accuracy in the entire brain and the GTV. Fully sampled T1-weighted MRI data from 73 pediatric patients treated for brain tumor or postoperative tumor bed were used to simulate 4-fold acceleration, which were split into the training (n=58) and validation (n=15) datasets.
Results: The adaptive DDPM showed significantly higher accuracy than TC-DiffRecon within GTV (PSNR, 30.1±1.9 vs 24.2±2.7) as well as over the entire brain (SSIM, 0.96±0.01 vs 0.91±0.02; PSNR, 28.5±1.5 vs 22.3±1.1). The tissue textures within the GTV were often sharper and more visually appealing with TC-DiffRecon, but their specific structures were not necessarily consistent with those in the fully sampled images.
Conclusion: The superior performance of the adaptive DDPM in GTV supports its potential for clinical use in RT planning and highlights the importance of conditioning a generative model with sufficient patient-specific information. Improperly conditioned models can generate fake details, which would mislead RT planning.