Room: AAPM ePoster Library
Purpose: beam CT (CBCT) Hounsfield units (HU) accuracy is one of the important parameter for accurate dose calculation in Adaptive Radiotherapy. The purpose of this study was to predict a synthetic CT (sCT) with correct HUs from the daily cone-beam CT (CBCT) using two machine learning models.
Methods: treated 17 Nasopharynx patients who had pair of planning CT (pCT) and first day CBCT images were included. All the pCT and CBCT images were rigidly aligned and resampled to voxel size of 1x1x3mm3. Two machine learning models were investigated for synthetic CT prediction with the hold-out validation approach (10 patients for training and 7 patients for testing)
First, a linear-mixed-effect (LME) model was trained with the pCT images as responses and CBCT images along with 6 additional features were used as predictors. The image features derived from CBCT, includes 2 Gabor-filtered images, 3 Sobel gradient images and an auto-threshold classified image (air-cavities, bone and soft-tissue).
Second, a convolutional neural network (CNN) was trained (2 hidden layers and 10 hidden neurons) for the same data.
Model prediction accuracy was evaluated using mean absolute error (MAE) of HU between the pCT and sCT images.
Results: average MAE between pCT and first day CBCT images within the external body contour was 203.2±32.7. The same between pCT and sCT images generated by LME model was 122.4±18.5. On the other hand the CNN model prediction resulted the MAE of 103.9±11.8. The training and prediction time (2.0GHz processor with 32GB RAM) were 7 minutes and <1 minute for LME and 370minutes and <1minute for CNN.
Conclusion: of predicting sCT from CBCT was explored using two machine learning approaches and the preliminary results were promising. CNN based sCT generation from daily CBCT resulted better HU accuracy than LME model.
Not Applicable / None Entered.