Purpose: Magnetic resonance imaging (MRI) has become an emerging tool in radiation therapy. It improves treatment accuracy over conventional CT-based radiotherapy due to superior soft-tissue contrast. While dose calculation is usually performed with CT images, MRI contrast can be critical for accurate tumor and organ delineation and segmentation. Therefore, we developed a supervised cross-domain adaptation based deep learning method to generate synthetic MR from CT datasets, facilitating tumor and organ auto-segmentation.
Methods: Current deep learning approach achieves reasonably accurate tumor segmentation from limited MRI data by leveraging large CT datasets. This method helps to preserve tumors on synthesized MRIs produced from CT images. We present a supervised cross-domain adaptation-based deep learning approach to generate synthetic MR from registered CT and T2-weighted MRI datasets. This approach was implemented following the steps: (1) performed image registration of CT and MRI on 15 lung cancer patients (IRB approved); (2) paired CT and MR data from 15 cancer patients was used for Unet training to generate training model for adaption; (3) a CT image from the 16th patient served as the test image for the training model, to generate synthetic MRI. The test CT image also served as the reference for image quality evaluation. Mutual information was calculated for the synthetic MRI and the test CT, and it was used as an evaluator of the trained model. This method was subsequently evaluated by comparing supervised cross-domain adaptation model with unsupervised cross-domain adaptation.
Results: Mutual information of the synthetic MRI and reference CT calculated from supervised cross-domain adaptation is 1.41; with unsupervised cross-domain adaptation it is 1.18, lower than the former.
Conclusion: The novel supervised cross-domain adaptation based deep learning approach showed promising results, achieving more accurate image features on synthetic MRIs, compared with unsupervised cross-domain adaptation based deep learning approach.