Room: AAPM ePoster Library
Purpose: This study proposes a novel image super-resolution method for proton dose distributions using a convolutional neural network (CNN).
Methods: First, 60 head and neck patient CT images with clinical target volume contours were collected from the TCIA database. These were divided into 6107 CT images (50 patients) for training of a CNN model, 674 (5) for validation, and 868 (5) for testing, respectively. Second, both low-resolution (4×4 mm2) and high-resolution (1×1 mm2) proton dose distributions were calculated for each patient from a random beam angle using matRad. Third, a DenseNet-based CNN was generated to predict a 256×256 high-resolution dose distribution by using a 64×64 low-resolution dose distribution and 256×256 high-resolution proton stopping power ratio image as inputs. End-to-end patch-based training was implemented until mean absolute dose error of the validation data was minimized. For evaluation, the CNN-calculated dose distributions of test data were compared with dose distributions upsampled by bicubic interpolation and the actual distributions.
Results: The CNN model calculates high-resolution proton dose distributions accurately. Averaged absolute dose error and its standard deviation over all test data were 1.36±5.13 cGyE for the CNN whereas 3.07±8.62 cGyE for the bicubic interpolation. Moreover, 3D Dice similarity coefficients between calculated and true distributions illustrate that CNN-calculated dose distributions are more accurate than bicubic-upsampled dose distributions. Relatively high dose errors were observed at distal fall-off dose edges. Computation time for calculating one dose distribution is around 6 milliseconds with a NVIDIA RTX 2080Ti GPU.
Conclusion: A novel image super-resolution method was established for proton dose distribution calculations using a CNN. This technique will be useful for generating detailed dose distributions from existing dose distributions or to accelerate dose calculation time. Our deep learning model will also be generalizable to extend super-resolution to 3D volumes.