基于粒子群算法与反向传播神经网络的呼吸运动预测研究

Respiratory Signals Prediction based on Particle Swarm Optimization and Back Propagation Neural Networks

在放射治疗过程中,呼吸运动会造成某些器官组织如肺、肝的靶区发生变化,从而降低放疗的效果,并且加大对正常组织器官的伤害。因此,在放疗过程中对靶区进行呼吸运动的实时估计是一项非常必要的工作。由于具备较好的非线性拟合能力,优化反向传播神经网络(BP-NN)已经被广泛应用于呼吸的预测,然而BP-NN容易陷入局部最优值。提出一种应用粒子群算法(PSO)优化BP-NN的方法减少陷入局部最优值的机率,提高呼吸运动预测的精度。首先,应用PSO算法寻找神经网络的最佳初始权值与阈值;然后,应用最优的初始权值与阈值建立神经网络(PSO-NN);最后,利用建立的PSO-NN网络进行呼吸预测。结果表明,11组肺癌病人呼吸运动预测实验对比结果表明,此算法(PSO-NN)相比单纯应用BP-NN算法的平均绝对误差由0.24减少到0.18(25%),互相关系数由0.82提高到0.86。所提出的算法可以有效地减少BP-NN陷入局部最优值的机率,提高预测的精度。

Objective, Respiratory motion may cause the change of some organs and tissues, such as lung and liver tumors in radiation therapy, which may influence the treatment effect and increase the damage to normal tissues and organs. Hence, it is an essential work to estimate the real-time movement of target in radiation therapy. Method, BP-NN has been widely used in respiratory motion prediction due to its superior non-linear fitting capability. However, BP-NN is easy to fall into local minimum. Results, In this study a novel method using PSO to optimize the BP-NN was proposed to avoid its drawbacks and improve prediction accuracy. Firstly, the PSO method was used to select optimal initial weights and thresholds of neural network. Then, the optimal initial weights and thresholds was utilized to establish artificial neural network (ANN). Finally, the established PSO-NN was used to predict respiratory signals.The preliminary results of 11 patients demonstrate that the mean absolute error reduced from 0.24 to 0.18(25%) and the coefficient correlation increased from 0.82 to 0.86. Conclusion, The proposed method (PSO-NN)could reduce the risk of BP-NN falling into local optimum and has the ability of improving the prediction accuracy of BP-NN method.