基于深度学习和字典学习构建多模态影像引导ESCC自适应放疗新策略

放化综合治疗是局部晚期ESCC重要治疗手段，但五年生存率仅有15-30%。治疗抗拒和食管疗中形变是疗效欠佳的可能原因。若能基于不同患者之间动态影像信息，早期预测ESCC放化疗敏感性及正常组织损伤，制定个体化的自适应放疗方案，对于提高局部晚期ESCC生存时间具有重要临床价值。本项目拟利用深度学习框架开发CT、CBCT和fMRI等不同模态影像之间形变配准算法，对ESCC放疗过程中解剖及功能变化进行综合评估，实现不同模态之间的优势互补；利用字典学习提取高区分度影像组学特征，评价肿瘤区域内部的生物异质性差异，解决ESCC放化抵抗和正常组织损伤的早期精准预测问题；在此基础上，结合靶区及正常组织剂量学变化，建立ESCC自适应放疗辅助决策模式，一方面预测自适应放疗计划再优化的最佳时机，另一方面利用放射生物学指标对原有放疗计划进行调整，最终为ESCC患者提供“量体裁衣”式的个体化治疗方案。

Radiochemotherapy is an important treatment for locally advanced esophageal squamous cell carcinoma (ESCC). However, the five-year survival rate remains only at 15-30%. Possible reasons for the poor efficacy can be treatment resistance and anatomic changes during radiotherapy. An individualized treatment regimen with adaptive radiotherapy may have a significant survival benefit for locally advanced ESCC patients, if treatment decisions for different patients can be made based on early prediction of the sensitivity of ESCC chemoradiotherapy and normal tissue damages using dynamic multi-modality image information. This project aims at developing algorithms for deformable registrations of various image modalities, such as CT, CBCT and fMRI, and using a deep learning framework, with anatomical and functional changes of ESCC radiotherapy assessed comprehensively using complementary information of multiple image modalities. Dictionary learning will be used to extract high-resolution radiomics features to evaluate biological heterogeneity in the tumor, along with early and accurate predictions for radiochemical resistance of ESCC and normal tissue damage. The dosimetric variations of the target and normal tissues will be incorporated to establish a decision-assisting model for ESCC adaptive radiotherapy, in order to help forecasting the best time to re-optimize the adaptive radiotherapy plans and adjusting the original care plans based on radiobiological indicators. As a result, ESCC patients will be offered “tailored” individualized treatment regimens.