Pre-training Model for Dose Distribution Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models for dose distribution prediction in radiation therapy are poorly generalized and struggle to adapt to different targets, requiring significant manual intervention and elongating the development time of radiotherapy plans.
Innovation Solution
A pre-training method and system for a dose distribution prediction model that adjusts based on multiple training tasks, using meta-learning techniques to generate a pre-training model that can quickly adapt to new tasks with a small number of samples, incorporating conditional generative adversarial networks for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a dose distribution prediction model is trained on a single specific target (e.g., nasopharyngeal site), then the model achieves high accuracy for that specific target, but the model has poor generalization capability and cannot adapt to different targets (e.g., prostate site)
Solution Approach 1:
The patent applies multi-functionality by training a single prediction model on multiple different targets (nasopharyngeal site, prostate site, lung site, etc.) simultaneously. The model learns universal features and patterns across different anatomical sites, enabling it to generalize to new targets without retraining. This resolves the contradiction by making the model adaptable to various targets while maintaining prediction accuracy through the unified training approach.
Solution Approach 2:
The patent segments the training process into two phases: pre-training on multiple different targets to learn universal features, and then fine-tuning on specific targets when needed. This segmentation allows the model to first acquire broad generalization capability and then specialize for specific applications, balancing both accuracy and adaptability.
2Manufacturing precision
If manual intervention is used to adjust plan parameters based on experience and feedback, then the quality of radiotherapy plans is improved, but the development time of radiotherapy plans is significantly elongated
Solution Approach 1:
The patent implements self-service by using the trained prediction model to automatically generate initial optimization targets and dose distribution predictions without requiring manual adjustment. The model learns from historical expert plans and automatically applies this knowledge to new cases, reducing both the time required and the dependency on manual expert intervention while maintaining plan quality.
Solution Approach 2:
The patent applies preliminary action by pre-training the model on extensive historical data and expert plans before actual use. This preliminary training phase captures expert knowledge and patterns, so that during actual plan development, the model can quickly provide accurate predictions without requiring time-consuming manual adjustments, thus reducing plan development time while preserving quality.
3Adaptability or versatility
If a prediction model is trained on multiple different targets, then the model achieves good generalization capability, but the model requires significant training data and computational resources
Solution Approach 1:
The patent merges training across multiple different targets into a unified training process. By combining data from nasopharyngeal, prostate, lung, and other sites into a single training framework, the model learns shared features and patterns that improve generalization. This merging approach efficiently utilizes available data across targets, reducing the total training data volume needed compared to training separate models for each target while maintaining strong generalization capability.
Data Source
AI summary
The present disclosure provides a pre-training method for a dose distribution prediction model, implemented on a device including at least one processor and at least one storage device. The method comprises: obtaining a plurality of training tasks; for each of the plurality of training tasks, obtaining one or more samples corresponding to the training task; and obtaining a pre-training model based on the samples corresponding to the plurality of training tasks. The pre-training model is configured to obtain a dose distribution prediction model for a target task by adjusting, based on one or more samples corresponding to the target task, the pre-training model. The dose distribution prediction model is configured to output predicted dose distribution information corresponding to the target task. For each of the samples of the plurality of tasks and the target task, a label of the sample includes labeled dose distribution information corresponding to the task.