放射多组学协同学习预测鼻咽癌自适应放疗触发机制

Radioactive multi-omics collaborative learning for adaptive radiation therapy eligibility prediction in nasopharyngeal carcinoma

针对传统的放射多组学(影像组学、剂量组学和轮廓组学)模型往往采用特征拼接的方式,容易忽略不同组学特定统计属性、产生过拟合的问题,提出了以一致性约束和自适应权重为核心构建的多组学协同学习算法(multi-omics collaborative learning,MOCL)。该算法采用一致性约束挖掘不同组学特征之间的互补模式,再通过香农熵自适应学习不同组学特征的权重,最后引入紧致度图来避免过拟合现象。通过将MOCL在311名鼻咽癌患者组成的临床影像数据上得到的实验结果与3种传统的机器学习算法以及2种多视角算法进行比较,结果表明MOCL在多组学协同学习上,具有一定的优势,能为鼻咽癌自适应放疗资格预测提供有价值的决策依据。

Traditional radiation omics models,including radiomics,dosiomics,and contouromics,typically adopt feature splicing,which tends to ignore the specific statistical attributes of different omics and therefore leads to overfitting.A multi-omics collaborative learning(MOCL)algorithm focused on consistency constraints and adaptive weights was proposed in the study to address this problem.The MOCL algorithm employs consistency constraints to explore complementary patterns among heterogeneous omics features and adaptively learns their weights using Shannon entropy while avoiding overfitting through compactness mapping.An experiment was conducted on the clinical imaging data of 311 patients with nasopharyngeal carcinoma using MOCL.The experimental result is compared with three traditional machine learning algorithms and two multiperspective algorithms.The results demonstrate that MOCL has certain advantages in collaborative learning of multi-omics and can provide a valuable prediction basis for adaptive radiotherapy qualification in the case of nasopharyngeal carcinoma.