大数据学习非规则放疗剂量分布预测研究

Existing knowledge-guided radiation treatment planning methods are commonly limited to the prediction of a few key parameters for the evaluation of plan quality. They fail to directly estimate irregular dose distributions although they are more clinically meaningful. To address this issue, this project explores big data analysis to model geometry and shape relationships between tumors and adjacent organs for the prediction of their dose distributions. Firstly, we study geometry and shape features that can accurately describe spatial relationship between tumors and adjacent organs of the current patient. The clustering algorithm is also studied to fuse these features from different domains. Fisher vector aggregation is a potential solution that clusters multi-domain features to find their correlations. Secondly, dictionary learning and sparse coding are investigated to search for a subset of feature components that can maximally represent the aggregated vector and reduce feature dimension within the foundation of big radiation data. The reduced feature vector is then exploited as the retrieval query to find relevant cases in the big radiation data using sparse coding method. Finally, group-wise image registration is explored to fuse dose distributions from the retrieved cases, and achieve dose reconstruction and prediction for the current patient. Incorporating the prediction of key parameters for radiation plan evaluation, the reconstructed dose distribution is further refined to improve the prediction accuracy. Through the above methods, with the development of big data analysis in the field of radiation therapy, this project is expected to provide a breakthrough approach that directly predicts irregular radiation dose distributions.

现有的放射治疗计划预测方法局限于计划评估参数建模，而不能预测对临床更为重要的非规则放疗剂量分布。针对这类问题，本课题开展大数据条件下基于肿瘤和相邻器官几何形状建模的非规则剂量分布预测研究。具体研究内容：1）研究对几何形状特征的精确描述，以及多模态特征的自动聚类算法，基于矢量化特征融合方法，有效地聚类肿瘤和相邻器官的几何形状特征，得到多模态特征的关联性；2）研究在大数据条件下字典学习和稀疏编码技术对多模态特征进行降维并提取有效子特征，实现对几何形状特征的检索表达，并精确搜索相关病例。3）研究运用群配准方法来融合相关病例的非规则剂量分布，以实现当前病人的非规则剂量分布重建及预测，并结合计划评估参数的预测结果，对重建结果局部调整以提高预测精度。本课题通过以上研究结合放疗领域大数据发展的趋势，可望为非规则放疗剂量分布预测的研究提供突破性发展的新路径。

现有的放射治疗计划预测方法局限于计划评估参数建模，而不能预测对临床更为重要的非规则放疗剂量分布。针对这类问题，本课题开展大数据条件下基于肿瘤和相邻器官几何形状建模的非规则剂量分布预测研究。具体研究内容：1）研究对几何形状特征的精确描述，以及多模态特征的自动聚类算法，基于矢量化特征融合方法，有效地聚类肿瘤和相邻器官的几何形状特征，得到多模态特征的关联性；2）研究在大数据条件下字典学习和稀疏编码技术对多模态特征进行降维并提取有效子特征，实现对几何形状特征的检索表达，并精确搜索相关病例。3）研究运用群配准方法来融合相关病例的非规则剂量分布，以实现当前病人的非规则剂量分布重建及预测，并结合计划评估参数的预测结果，对重建结果局部调整以提高预测精度。本课题通过以上研究结合放疗领域大数据发展的趋势，可望为非规则放疗剂量分布预测的研究提供突破性发展的新路径。