Room: Exhibit Hall
Purpose: Modeling complex systems with deep learning is a challenging and computationally intensive task since it requires a large data set and a deep and complex neural network. However, the complex behavior can sometimes be approximately described by an analytical model. We hypothesize that, by incorporating this easier-to-obtain prior knowledge into deep learning models, the problem becomes sparse and thus the required data size and model complexity can be greatly reduced and the computational efficiency can be improved.
Methods: To test this hypothesis, we use radiotherapy dose calculation as a toy model, although deep learning is not really necessary to solve the problem. To accurately model the dose calculation from IMRT fluence maps to 3D dose distributions is extremely challenging for deep learning. Our work converts this challenging task to a relatively easy task by incorporating the prior knowledge which is the broad beam ray tracing (bb) result for each beam. The desired 3D dose distribution was calculated using the model-based collapsed cone convolution/superposition algorithm (cccs) and used for training and testing. We employed a Convolutional Neural Network (CNN) to model the residual sparse and localized relationship between ray tracing results and the convolution/superposition results. The CNN model was trained on 71 prostate IMRT patients, each with seven 6MV beams, with 5-fold cross validation. A separate data set of 8 patients was used for testing.
Results: On average, the PTV coverage and max dose was predicted within 1.2%(D95), 0.7%(D98), 1.0%(D99), and 5.3%(Dmax), an improvement over the broad beam ray tracing results of 3.8%(D95), 4.6%(D98), 5.7%(D99), and 5.3%(Dmax).
Conclusion: The trained CNN achieved accurate prediction of 3D dose distributions. This simple example supports our hypothesis that the incorporation of prior knowledge can greatly reduce the requirement on data size and model complexity when solving complex problems.