Purpose: A major challenge to the clinical implementation of MRI-only simulation is validation of the integrity of synthetic CT (sCT) images. In this study, we produced criterion for clinical acceptability of sCT image generation by correlating image similarity metrics to dosimetry and alignment errors using only MRI-simulation and their corresponding sCT images.
Methods: MRI-simulation images from 5 pelvic cancer patients were included in this study. Bulk-density (BD) and deep-learning (DL) model-based sCT images were generated using two commercially available sCT methods. Criteria for clinical acceptability was established using MRI/sCT image correlation scores and dosimetry/alignment errors after adding artifacts into each sCT. To emulate artifacts commonly seen in MRI-based sCT generation, bone and air artifacts of varying volumes were inserted into different parts of sCT images. Over 400 corrupted sCT images were generated for each patient containing bone/air artifacts. Mutual-information (MI) and structural-similarity-index-metric (SSIM) scores were calculated as proxies for the MRI/sCT image similarity. A 5-field intensity-modulated-radiation-therapy (IMRT) prostate plan was created using the original sCT image and doses were recalculated onto corresponding corrupted sCTs. sCT vs corrupted-sCT gamma indices and mean planned dose-differences were reported. MRI/sCT image translational errors are reported for each corrupted-sCT.
Results: SSIM and normalized-MI scores that resulted in a planned-dose-difference below 2% and local gamma-index (2%/2mm/10% threshold) below 98% were 0.996 (bone-artifacts) (BD-method) and 0.995 (bone-artifacts) (DL-method), and 0.994 (bone/air artifacts) (BD-method) and 0.993 (bone/air artifacts) (DL-method), respectively. Root-mean-square translational alignment differences for all tested corrupted-sCT images were below 2mm.
Conclusion: This work has demonstrated the proof-of-concept that image similarity metrics can be used to provide clinical endpoint-based metrics to allow for evaluation of sCT images produced using the MRI-only workflow. Future work will involve including additional patient statistics, addition of more sCT generation methods and the inclusion of other clinical sites such as head/neck.
Not Applicable / None Entered.