基于多模态功能影像组学及ctDNA预测食管癌新辅助放化疗疗效的研究

Establishing the model for predicting pCR (pathologic complete response) will guide neoadjuvant treatment decisions and personalized treatment of esophageal cancer. Reported models yield unsatisfactory results due to low accuracy and could not be applied widely. The combination of oncogenic mutation information and functional radiomics integrates metabolic, microstructure and mutation information, and will dramatically improve the predictive ability. We have identified that treatment-induced change in ADC during the first 2 weeks of nCRT for esophageal cancer seemed highly predictive of pCR. We have recently focused on the study of radiomics and machine learning, and successfully established the radiomics model based on 48 MRI features. The combination of MRI-DWI and PET-CT has not been used in the study of radiomics, and non-invasive ctDNA detection was also not integrated into radiomics model yet. This study will prospectively collect the tissue and blood samples of locally advanced esophageal cancer patients receiving neoadjuvant chemoradiotherapy (nCRT) and surgery at different time points. We will use optimized machine learning method to establish pCR prediction model based on ctDNA mutation burden and CNV (copy number variants) in combination of multiple MRI-DWI and PET-CT features, and we will compare the predictive efficacy of different models and affirm the best prediction timeline. We will further validate the model in another group of esophageal cancer patients and establish a model for survival prediction by longer follow up time. This model might possibly guide the personalized treatment for esophageal squamous cell carcinoma patients in China.

建立预测放化疗后pCR的模型对食管癌个体化治疗十分重要。文献报道的模型准确性低，难以推广。联合癌突变信息的功能影像组学可有效整合代谢、微结构和突变这三类数据，提升预测pCR效能。我们发现放疗后2周MRI-DWI的ADC值升高可预测食管癌pCR。我们近年一直从事影像组学和机器学习的研究(发表3篇SCI)，建立了基于48个MRI特征的影像组学模型。联合MRI-DWI和PET-CT的"多模态功能影像”尚未用于影像组学研究，反映癌突变的ctDNA无创检测也未整合入影像组学模型。本研究将在不同时间点收集食管癌患者组织和血液标本，用靶向测序检测ctDNA突变负荷及基因拷贝数，采集PET-CT和MRI-DWI特征参数，优化机器学习建模策略，建立预测pCR的模型，比较不同模型效能，确立最佳预测时间点。在另一组患者独立验证，随访后建立生存预测模型。该模型有望为临床决策提供新依据，指导食管癌的个体化治疗。

食管癌是最常见的恶性肿瘤之一，是全球癌症相关死亡的第六大常见原因。食管鳞状细胞癌是中国最常见的食管癌亚型。目前，根治性放化疗已成为目前不可切除的局部晚期食管癌的标准治疗方法。尽管放射治疗技术发展迅速，但不可切除的食管癌的预后仍然令人失望。超过50% 的食管癌患者最终在根治性放化疗后出现局部复发，5年总生存率为15-25%。预测局部复发有助于医生为具有不同风险的食管癌患者提供个性化治疗，并且对于复发风险高的患者应尽早实施适当的治疗策略。影像组学是一种通过高通量提取定量特征将医学图像转换为数据的工具，在癌症诊断，预后和治疗反应等预测很有前途。利用影像组学联合深度学习模型预测食管癌局部复发的多中心研究还未见报道。本项目在西京医院回顾性收集的302例食管鳞癌患者，随机划分得到201人的训练集和101人的内部验证集，从天津和山东肿瘤医院收集的70例患者作为外部验证集。对每位患者，采集治疗前的增强CT和临床资料，提取radiomic signature和deep-learning signature。Radiomic signature和deep-learning signature，联合预后临床危险因素，建立并验证基于治疗前增强CT的融合列线图预测模型，预测局部无复发生存期。基于治疗前增强CT的radiomic signature和deep-learning signature可以将食管鳞癌患者风险分层，分为高危组和低危组，对预测食管鳞癌的局部无复发生存期和总生存期具有重要价值。将radiomic signature、deep-learning signature与临床数据相结合建立的融合列线图优于使用三个预测因子的子集模型，在训练集、内部验证集和外部验证集中一致性指数分别为0.82、0.78和0.76。总之，本项目开发并验证了基于增强CT的影像组学列线图，该模型结合了radiomic signature，deep-learning signature和临床因素，并在多中心数据评估其对接受根治性放化疗的食管鳞癌患者的复发风险的预测性能。该列线图作可作为预测根治性放化疗的食管鳞癌患者预后的有力工具，指导食管癌的个体化治疗。