Cross-technique transfer learning to predict the dose distribution for radiotherapy planning based on a limited sample size.
Accurate dose prediction is challenged by the lack of available training samples and the rapid evolution of radiotherapy techniques.
A cross-technique transfer learning strategy was developed to predict the dose distribution for radiotherapy planning using limited training samples.
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). The Res-U Net was selected as the base deep learning network. Cross-technique models were pretrained on the IMRT dataset and subsequently fine-tuned on VMAT data using limited samples (five and seven cases). Independent models were trained from scratch using the same limited samples, while a standard model trained on the full VMAT training set served as the reference. Model performance was evaluated on a test set using metrics including the dose-volume histogram (DVH), voxel-based mean absolute error (MAE), and the Dice similarity coefficient (DSC) of the isodose volume.
The cross-technique models exhibited clinically acceptable performance with only five training samples and were comparable to the standard model (MAE deviation: 0.15%, p > 0.01 after Bonferroni correction; DSC deviation: 0.11%-0.72%). Performance improved further with seven training samples (MAE deviation: 0.05%, p > 0.01; DSC deviation: 0.02%-0.40%). However, the independent models trained with five or seven samples showed significantly inferior performance (five samples: MAE deviation: 1.14%, p < 0.01, DSC deviation: 0.98%-2.48%; seven samples: MAE deviation: 0.50%, p < 0.01, DSC deviation: 0.48%-1.05%).
The cross-technique models accurately and reliably predicted the dose distribution for a new radiotherapy technique using a limited sample size.
A cross-technique transfer learning strategy was developed to predict the dose distribution for radiotherapy planning using limited training samples.
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). The Res-U Net was selected as the base deep learning network. Cross-technique models were pretrained on the IMRT dataset and subsequently fine-tuned on VMAT data using limited samples (five and seven cases). Independent models were trained from scratch using the same limited samples, while a standard model trained on the full VMAT training set served as the reference. Model performance was evaluated on a test set using metrics including the dose-volume histogram (DVH), voxel-based mean absolute error (MAE), and the Dice similarity coefficient (DSC) of the isodose volume.
The cross-technique models exhibited clinically acceptable performance with only five training samples and were comparable to the standard model (MAE deviation: 0.15%, p > 0.01 after Bonferroni correction; DSC deviation: 0.11%-0.72%). Performance improved further with seven training samples (MAE deviation: 0.05%, p > 0.01; DSC deviation: 0.02%-0.40%). However, the independent models trained with five or seven samples showed significantly inferior performance (five samples: MAE deviation: 1.14%, p < 0.01, DSC deviation: 0.98%-2.48%; seven samples: MAE deviation: 0.50%, p < 0.01, DSC deviation: 0.48%-1.05%).
The cross-technique models accurately and reliably predicted the dose distribution for a new radiotherapy technique using a limited sample size.