Purpose: We predicted the dose distributions from segmented anatomy for radiation treatment planning in our previous paper. However, the density information derived from CT images and the distance to tumor target can affect the prediction accuracy. In this study, we employed geometric and imaging information with segmented anatomy for deep-learning based dose distributions prediction and compared the accuracy of dose distributions predictions with different inputs to find a method to get better accuracy.
Methods: Ninety-nine cases of nasopharyngeal cancer were included in the study. Eighty-nine cases were chosen randomly as the training set and the remaining as the test set. The prediction model was based on the ResNet-101. We compared the dose predictions with four kinds of inputs, i.e. images of CT, segmented anatomy, anatomy with distance to planning target volume (ADTP), and the combination of CT, anatomy and distance to planning target volume (Com). The accuracy of predicted dose distributions was evaluated against the corresponding dose as used in the clinic. The voxel-based mean absolute error (MAE) was used as the evaluation index.
Results: The MAE of the whole volume was 8.00Â±1.68%, 5.31Â±0.57%, 5.41Â±0.71%, and 4.99Â±0.46% for the prediction with CT, anatomy, ADTP, and Com, respectively. The method with input of combination had significantly better results; with p value of 0.0002, 0.0078, and 0.0229 compared the methods with other three inputs, respectively.
Conclusion: The quantitative results show that the combination of images of CT, segmented anatomy, and distance to planning target volume yields best result in dose distribution prediction. This is mainly due to all the information is closely relative to the treatment plan. This study provides a basis for designing methods to obtain high precision dose prediction.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by Beijing Hope Run Special Fund of Cancer Foundtion of China (LC2018A14), the National Natural Science Foundation of China (11875320,11605291) and the National Key Projects of Research and Development of China (2017YFC0107500). The authors report no conflicts of interest with this study.
Not Applicable / None Entered.