Dosimetric Evaluation of Multi-sequence MR Image-based Synthetic CT Generation for Head and Neck MR-only Radiotherapy

Abstract

MR-only radiotherapy has attracted a great interest in head and neck to eliminate the inter-modality registration uncertainties and reduce the additional cost from a CT scan. Nevertheless, synthetic CT (sCT) should be generated to provide electron density information for both dose calculation and patient positioning purpose in MR-only workflow. Specific sequence or multi-sequence MR images are usually scanned to facilitate physicians to diagnose lesions and delineate target and normal organ structures. However, no studies have shown which MR sequences or combination of several sequences are suitable to be chose from the view of sCT generation. This work therefore aims to investigate the effect of different MR sequences on the accuracy of sCT generation and to proceed dosimetric evaluation on sCTs from different MR sequences in head and neck. Four kinds of MR sequence images (T1, T2, T1C and T1DixonC) and CT images were collected from 45 patients with nasopharyngeal carcinoma. 30 patients were randomly selected to train five mapping models (four single-channel models and one 4-channel model, each channel used an MR sequence image as input) with conditional generative adversarial network (cGAN). The remaining 15 patients in the testing set were used for evaluation. All predicted sCTs were then imported into TPS for dose calculation. Both sCT image quality and dosimetric accuracy were evaluated by various metrics. Among all the five models, the 4-channel model achieves highest accuracy in both sCT image quality and dosimetric accuracy than any single-channel model. Among the four single-channel models, the T1-weighted MR model has the best performance. The dose comparison of the five models in target and OARs show that the MAE values of PTVs are all less than 1% of prescription dose, and the MAE values of parotids, lymph nodes, brainstem and spinal cord are all less than 0.6%. The differences for dose-volume criteria of PTVs and OARs from all five models are all within -1.5%∼1.0%. Five sCT generation models based on multi-sequence MR images were successfully established with cGAN for head and neck. Both the dosimetric evaluation and sCT comparison show that the performance of 4-channel model and T1 model is slightly better than other models, while the dosimetric difference is relatively small. All dosimetric evaluations suggest that the sCTs generated by all sequence models can be applied to the process of MR-only radiotherapy from the dosimetric point.Tabled 1Abstract 2652; TablesCTDoseMAE(HU)SSIMPSNRMAE (%)2mm/2% (%)T175.22±11.480.84±0.0228.85±1.110.29±0.1099.25±0.22T1C80.03±10.870.82±0.0328.40±1.070.29±0.1099.20±0.26T1DIXONC86.27±10.820.79±0.0327.69±0.920.31±0.1099.10±0.26T286.95±10.780.78±0.0327.52±0.940.31±0.1199.09±0.254-channel69.98±12.020.85±0.0329.39±1.290.28±0.1099.32±0.24 Open table in a new tab