.png){kind=icon}
Return
Multi-stage framework with difficulty-aware learning for progressive dose prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.bspc.2022.104541.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Purpose: In this study, we are concerned with overcoming the limitation that most existing methods learn the dose prediction distribution without taking into account hard-to-predict regions. Methods: This paper presents a novel multi-stage dose prediction framework based on difficulty-aware learning to progressively predict 3D radiation dose distribution. We first stack multiple sub-models employing an encoder-decoder architecture to construct a multi-stage framework that decomposes the challenging prediction task into a few easy-to-learn sub-tasks for improving dose prediction. Moreover, a difficulty-aware learning strategy with voxel-level attention is developed to effectively handle hard-to-predict voxels or regions. Finally, the performance of our framework is compared to that of existing related models while being evaluated by voxel-based mean absolute error (MAE) and clinical-related dosimetric metrics. Results: We evaluated the proposed method on the head-and-neck datasets from the 2020 AAPM OpenKBP challenge. Our framework achieved a dose score of 2.367 and a DVH score of 1.378, which outperformed the other existing single-stage and cascade dose prediction models. The average prediction error of the dose CI and dose HI in the test dataset of 56 Gy, 63 Gy, and 70 Gy were quite similar to the ground-truth plans (ranging from 0.01 to 0.02). For the OARs, the average error in predicting the Dmax value was 2.66 +/- 2.88 Gy (brainstem), 1.94 +/- 1.43 Gy (spinal cord), and 1.39 +/- 1.67 Gy (mandible), whereas the values for the Dmean value were 1.10 +/- 1.07 Gy (left parotid), 1.13 +/- 1.03 Gy (right parotid), 1.81 +/- 1.77 Gy (esophagus), and 1.49 +/- 1.66 Gy (larynx).Conclusion: Both quantitative and qualitative results demonstrated the potential of our framework and its superior performance compared to the state-of-the-art methods in the dose prediction task, which thus proved to be a promising tool for predicting the dose distribution in computer-aided radiotherapy planning.
Keywords:
Radiation therapy
Dose prediction
Convolutional neural network
Difficulty-aware learning
Journal
IF:
4.9
Papers:
1.0W
Citations:
2.4W
Organization
Cited Papers
Multi-constraint generative adversarial network for dose prediction in radiotherapy.png)
MEDICAL IMAGE ANALYSIS
MEDICAL IMAGE ANALYSISIF11.8
PREAI
ShareSaveGeneration of hydroxyl radicals by urban suspended particulate air matter. The role of iron ions
PREAI
ShareSaveShareSaveA deep learning method for prediction of three-dimensional dose distribution of helical tomotherapyMEDICAL PHYSICS
MEDICAL PHYSICSIF3.2
PREAI
ShareSaveDifficulty-aware hierarchical convolutional neural networks for deformable registration of brain MR imagesMEDICAL IMAGE ANALYSIS
OAAI
MEDICAL IMAGE ANALYSISIF11.8
OAAIShareSaveTechnical Note: Dose prediction for head and neck radiotherapy using a three-dimensional dense dilated U-net architectureMEDICAL PHYSICSOAAI
MEDICAL PHYSICSIF3.2
OAAIShareSaveDose prediction using a deep neural network for accelerated planning of rectal cancer radiotherapy
PREAI
ShareSave
View more