Author: Benito De Celis Alonso, Braian Adair Maldonado Luna, Gerardo Uriel Perez Rojas, René Eduardo Rodríguez-Pérez, Kamal Singhrao 👨🔬
Affiliation: Department of Radiation Oncology, Brigham and Women's Hospital, Dana Farber Cancer Institute, Harvard Medical School, Faculty of Physics and Mathematics, Benemérita Universidad Autónoma de Puebla 🌍
Purpose: Artificial Intelligence (AI)-generated synthetic CT (sCT) images can be used to provide electron densities for dose calculation for online adaptive MRI-guided stereotactic body radiotherapy (SBRT). SCT model training uses deformable image registration (DIR) with MRI/CT image pairs which can result differences in positions of gas-cavities and bone structures. This can result in sCT image feature and dosimetric errors especially at gas-cavity interfaces when using sCT for MRI-guided SBRT. Here, we propose using a DIR-independent multi-stage tissue-label mediated sCT generation method to eliminate MRI/CT misregistration errors during sCT training.
Methods: SCT images were generated using a two-stage generative-AI model, 1) conversion of MRI to label map and 2) conversion of label map to an sCT image. Model preprocessing involved creation of tissue specific labels including gas-cavities, soft tissue, bone etc. using both manual and AI-based autocontouring. Labels were created independently for tissue features in MRI and CT. Both stages utilized default hyperparameters, and training was conducted using 4-fold cross-validation with paired images from 12 pelvic-cancer patients. MRI-label model was generated using CycleGAN and pix2pix with 100 epochs of training and model accuracy was evaluated using the Dice coefficient. Label-sCT model was evaluated using mean-absolute-error (MAE). Both metrics were applied to the anatomy as a whole and to tissue specific labels created.
Results: The pix2pix model achieved a Dice value of 0.98±0.01, while CycleGAN had a value of 0.91±0.04. CycleGAN also outperformed pix2pix in bone reconstruction, with DICE values of 0.44±0.01 compared to 0.31±0.1. In the second stage, the MAE values for pix2pix across all anatomy were 80±16HU, while for CycleGAN they were 82±20HU.
Conclusion: This study shows that incorporating tissue label-mediated sCT images can improve the accuracy of sCT-generating models. Future work will focus on incorporating online adaptive treatment planning studies to evaluate the dosimetric improvement of using this method.