Screen: 29
Xiaohong Wang, MD
beijing
Accurate dose prediction is challenged by the lack of available training samples and the constant emergence of new radiotherapies. A cross-technique transfer learning strategy was developed to predict the dose distribution for radiotherapy planning based on limited samples.
Materials/
Methods:
Data were collected from 154 patients with nasopharyngeal carcinoma: 60 treated with intensity-modulated radiotherapy (IMRT) and 94 treated with volumetric modulated arc therapy (VMAT). ResNet101 was selected as main part of the basic deep learning network. Cross-technique models were trained on the IMRT data and were then used to predict the dose distribution for VMAT from limited samples (5 and 7). Independent models were trained using the same limited samples. The model performances were evaluated by using the test set based on parameters including the dose–volume histogram (DVH), voxel-based mean absolute error (MAE), and dice similarity coefficient (DSC) of the isodose volume. A standard model fully trained on the VMAT data was used as the ground truth.
Results:
The cross-technique models performed well with only five samples and comparably to the standard model (MAE deviation: 0.15, DSC deviation: 0.11–0.73, P > 0.05). The performance improved further with seven samples (MAE deviation: 0.05, DSC deviation: 0.02–0.40, P > 0.05). The independent models performed significantly worse with five samples (MAE deviation: 1.14, DSC deviation: 0.98–2.48, p < 0.05) and seven samples (MAE deviation: 0.5, DSC deviation: 0.50–1.05, p < 0.05). The DVH predicted by the cross-technique models closely agreed with that of the standard model.
Conclusion:
The cross-technique models accurately and precisely predicted the dose distribution for a new radiotherapy from a limited sample size. The proposed strategy efficiently eliminates problems due to differences among techniques and data shortages to promote the swift implementation of novel radiotherapies.