肿瘤放疗并发症概率预测模型参数拟合方法

Parameter fitting method of NTCP predictive model in radiation oncology

为了建立具有群体特异性的肿瘤放疗NTCP预测模型,提出了一种模型参数拟合方法.首先,基于NTCP模型的特点构建最大似然函数;然后,分别采用确定性优化方法和随机性优化方法对最大似然函数进行优化,分析优化过程的时间成本及优化结果,探讨用于拟合NTCP模型参数的最优方法.实验结果表明,用于拟合NTCP模型参数的最大似然函数是非凸的,存在局部最优解;遗传算法是一种最稳定的最大似然函数优化方法,其运行时间比模拟退火算法短,而且可以在每次优化结束后给出全局最优解,以作为NTCP模型参数.所提方法可以帮助肿瘤放疗工作者在临床随访数据的基础上建立具有群体特异性的放疗并发症预测模型.

To establish the population specific NTCP（ normal tissue complication probability） prediction model in radiation oncology,a parameter fitting method is proposed. First,the maximum likelihood function is constructed based on the characteristic of the NTCP model. Then,the deterministic optimization method and the stochastic optimization method are used to optimize the maximum likelihood function,respectively. The time cost and the optimization results are analyzed to find the better method for fitting the parameters of the NTCP model. The experimental results show that the maximum likelihood function for fitting the parameters of the NTCP model is non-convex,indicating that there exist the local optimal solutions. The genetic algorithm is the most stable optimization algorithm for fitting the NTCP model,and the running time is less than that of the simulated annealing algorithm. In this algorithm,the global optimal solutions,which are regarded as the parameters of the NTCP model,can be obtained after each optimization. The proposed method can help the researchers in radiation oncology establish the population specific NTCP predictive models based on clinical follow-ups.