Exhibit Hall | Forum 2
Purpose: In order to optimize the sequence for acquisition time and image quality, the MR images for adaptive planning are often truncated in some boundary regions such as shoulders in head-and-neck. To address missing structures which limit MR-based dose calculation, we modified the cycle-consistent Generative Adversarial Network (cycleGAN) to generate synthetic CT (sCT) with full anatomy from such truncated MRIs.
Methods: We modified cycleGAN by adding one extra input channel and constructing two loss functions to ensure anatomy compensation and preserve structure similarity for sCT generation. Two pretrained UNet models were used to contour body volumes from sCT and real CT. A truncation loss based on dice similarity coefficient penalized the difference in body volumes. A structure-consistency loss based on modality independent neighborhood descriptor (MIND) was introduced between synthetic images and input images to maintain structure similarity. Our model was trained using 28 head-and-neck patients treated with an MR-Linac using simulation CT, truncated T1-weighted MRI, and body contours for each patient. Our model was tested using 8 patients and compared with conventional cycleGAN using the same training and testing dataset.
Results: Our model more accurately compensated the missing shoulder on truncated MRI than state-of-the-art cycleGAN. Also, our model created more accurate MR-consistent bony structures than cycleGAN. The mean absolute errors (MAEs) between our sCT and real CT were 72.1±16.3HU, 108.2±11.8HU, and 106.0±11.6HU within truncated regions, untruncated regions, and whole-body volume, respectively, outperforming the cycleGAN with 442.8±100.0HU, 148.9±15.0HU, and 169.3±18.0HU, respectively. Our model had statistically significant improvement over cycleGAN in MAEs, peak single to noise ratio (PSNR), and structural similarity index (SSIM) evaluation metrics (student t-tests with p<0.05).
Conclusion: We developed a novel cycleGAN based model for sCT generation from truncated MRI. Our model has demonstrated better accuracy in sCT generation for both truncated and untruncated regions than conventional cycleGAN.