2034 - PTU: A Pyramid Transformer Learning Framework with Low-Rank Approximation for Voxel-Level Dose Prediction of Radiotherapy
Presenter(s)
L. Chen, Z. Wang, T. Zhang, W. Wang, X. Sun, J. Duan, Y. Gao, Z. An, and L. N. Zhao; Department of Radiation Oncology, Xijing Hospital, Fourth Military Medical University, Xi'an, Shaanxi, China
Purpose/Objective(s): A novel Pyramid Transformer based U-net (PTU) model is proposed to accelerate the Intensity-modulated radiation therapy (IMRT) planning for head and neck tumors while more precisely predicting the voxel-level dose distribution.
Materials/Methods: The proposed end-to-end PTU model employs the pyramid Transformer to capture multi-scale features, and utilizes residual convolution blocks to extract shape-aware features which compensate the loss of spatial information due to the down-sampling in the pyramid structure. Moreover, online data augmentation is used to improve the prediction performance of the PTU. However, training and deploying the PTU with standard transformer can be prohibitively costly as the memory and computational complexity is quadratic in the inputs. To speed up the training and deploying of the PTU, we demonstrate low-rank characteristics of the weights of the PTU model and propose a model compression method that prunes the incoherent and non-expressive weights and approximates the remaining weights by a low-rank matrix. Pruning enhances the diversity of low-rank approximations, and low-rank approximation prevents pruning from losing too many expressive neurons. The PTU model is trained and evaluated on the dataset of the 2020 OpenKBP Challenge and its prediction performance is compared with three previous dose prediction models, C3D, TrDosePred, and TSNet.
Results: The predicted dose distributions of our PTU model are closest to the original clinical dose distribution. The PTU model achieves the dose score of 2.32 Gy and the DVH score of 1.34 Gy on the test dataset, outperforming the other three models by 5%-17%. Compared with other three models, the number of parameters, the training and inferring time of the PTU model are the least.
Conclusion: The quantitative results demonstrate that the proposed PTU model achieves more accurate voxel-level dose prediction. The PTU model has a great potential to improve the quality and efficiency of radiotherapy planning.