Author: Mingli Chen, Xuejun Gu, Hao Jiang, Mahdieh Kazemimoghadam, Weiguo Lu, Qingying Wang, Kangning Zhang 👨🔬
Affiliation: Medical Artificial Intelligence and Automation (MAIA) Lab, Department of Radiation Oncology, UT Southwestern Medical Center, UT Southwestern Medical Center, Department of Radiation Oncology, Stanford University School of Medicine 🌍
Purpose:
PET is used in radiotherapy workflows for accurate target delineation. However, a separate CT scan is typically required for attenuation correction in PET imaging and for registering PET-delineated targets to the planning CT. The requirement of separate CT imaging increases radiation exposure risks, particularly for vulnerable populations (e.g., pediatric, pregnant patients). Additionally, separate scans often result in misalignment due to patient motion and require complex image registration. This study proposes a deep learning-based (DL) method to generate synthetic CT (sCT) directly from PET images, reducing radiation exposure, resolving misalignment challenges, and eliminating the need for image registration, thereby enhancing the safety, efficiency, and cost-effectiveness of PET imaging.
Methods: Our model was built upon the 3D nnUNet framework, with the original loss function replaced by the Huber loss to address voxel-level intensity differences critical for PET-to-CT conversion. The dataset used in this study was obtained from the HECKTOR 2022 challenge, consisted of 524 paired PET and CT images of head and neck region. Of these, 422 pairs were used for training and validation, and 102 pairs for testing. During training, the model was trained to predict CT images using PET images as input. During testing, the model directly generated sCT images from PET inputs.
Results:
The results demonstrated strong agreement between sCT and CT in replicating Hounsfield Unit (HU) profiles and anatomical structures. sCT accurately preserved intensity values across soft tissues, air, and bone, with minor smoothing at sharp gradients. Additionally, sCT effectively mitigated metal-induced artifacts present in CT, reducing streaking and preserving structural details.
Conclusion:
We developed a DL method for sCT generation in PET imaging that aims to reduce radiation exposure, mitigate misalignment challenges, and is promising to offer safer and more efficient alternative to conventional PET/CT workflows, holding the potential to enable PET-only radiotherapy treatment planning.