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:
Converting MR images to synthetic CT (MR2sCT) is highly desirable as it streamlines the radiotherapy treatment planning workflow. This approach leverages the superior soft tissue visibility of MR for segmentation and the electron density information of synthetic CT for dose calculation and patient positioning, eliminating the need for CT imaging or MR-CT registration. We demonstrate the adaptation of 3D nnU-Net, with minimal code modification, by leveraging its self-configurable framework for the MR2sCT task.
Methods:
The adaptation of nnU-Net for MR2sCT involved only a modification to the loss function, replacing it with Huber loss to handle outputs in the continuous space. We trained and tested three models—T1w2sCT, T2w2sCT, and T2Flair2sCT—across various treatment sites and image sizes, all with a fixed resolution of 2 mm. The training/testing sample sizes for the three models were 125/29, 564/106, and 28/6, respectively, using registered MR-CT pairs. Couch removal was applied during preprocessing to mitigate artifacts and hallucinations in the predictions. The accuracy of the synthetic CT (sCT) was evaluated by comparing it to the registered CT in terms of Hounsfield Units (HU), dose calculation (T1w2sCT and T2Flair2sCT), and digitally reconstructed radiographs (DRR).
Results:
The overall average mean absolute error (MAE) was 73.85 HU, with higher MAE observed in brain and lung cases, due to registration misalignment rather than model performance. The average gamma passing rates for dose comparison using the 2%/2mm and 1%/1mm criteria were 98.6% and 93.77%, respectively, demonstrating accuracy of sCT for dose calculation. DRRs of sCT and CT highlighted misalignment in certain cases.
Conclusion:
Despite misalignment in registered MR-CT pairs and recognizing limitations of MAE in accuracy evaluation, the models generated accurate sCT for treatment planning, as most paired regions were aligned. With minimal coding required, this approach significantly accelerates model development and clinical implementation.