Author: Laura I. Cervino, Wendy B. Harris, Paulo Quintero, Hao Zhang 👨🔬
Affiliation: Department of Medical Physics, Memorial Sloan Kettering Cancer Center 🌍
Purpose: To evaluate the impact of the prediction uncertainty from CBCT-based synthetic CT (sCT) generation in abdominal adaptive radiotherapy.
Methods: CT and CBCT images from 65 abdominal patients were retrospectively retrieved to train three AI-based models to generate sCT images (M1: Unet, M2: Bayes-Unet, M3: cycle-GAN). To promote controlled variations in each model’s prediction uncertainty, two datasets (D1, D2) were assembled using D1: Rigid Image Registration (RIR) and D2: Deformable Image Registration (DIR) between CT and CBCT images. For every testing-dataset-input, the uncertainty maps for M1 and M3 were calculated using the ensemble-method by computing the voxel-wise standard deviation of 10 predictions from 10 independently trained sub-models. For M2, the uncertainty was the root squared variation from 100 iterations of the MC-DropConnect method. Treatment plans were optimized and calculated on ground-truth-CT and recalculated to their respective sCTs to evaluate the effect of uncertainty on dose calculations. Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Structural Similarity Index Measure (SSIM) were used to evaluate sCT image quality. Dose-Volume-Histogram (DVH) similarity was verified with t-test, and uncertainty maps correlation between Gamma Index (GI), MAE, RMSE, and SSIM was calculated with Spearman’s coefficient (Sp).
Results: For the three models {M1-M2-M3} trained with D1 and D2, the MAE were {50.9±13.3} and {40.9±11.5}, respectively, the RMSE were {68.3±13.5} and {62.2±10.7}, respectively, and the SSIM were {0.89±0.05} and {0.94±0.05}, respectively. No significant differences between DVH (p>0.005) were found. Correlations between uncertainty-maps and GI, MAE, RMSE, SSIM were statistically significant (Sp=0.84), moderate (Sp=0.55), moderate (Sp=0.58), and weak (Sp=0.38), respectively. The same evaluation considering only the PTV presented no correlation with any metric.
Conclusion: Uncertainty maps should be included in adaptive radiotherapy workflows to aid clinical decisions and increase the reliability of the synthesized images in terms of HU integrity and dose calculation accuracy.