面向医学检验数据分析的增强深度学习预测模型

Predictive model of reinforcement depth learning for medical laboratory data analysis

为了提升疾病早期筛查的准确性,文中基于临床检验数值、CT图像等数据对计算机辅助诊疗技术进行了研究。由于医疗图像存在伪影严重和边缘模糊的问题,若采用传统的深度学习算法(DL)需要增加网络深度,这对计算机的算力与训练样本数量均提出了更高的要求。为了克服上述问题,文中在传统算法的基础上引入了增强学习算法(RL),通过将两个算法进行融合改进,进而提出了一种DCQN网络。其利用多层非线性卷积网络,组合并提取图像的低阶特征,同时借助RL中的智能体(Agent)概念,不断累积环境迭代过程中的惩戒值,从而获得识别医学图像的最优路径。在某开放肺结节数据集上进行的仿真实验结果表明,DCQN网络在进行图像分割时,综合识别精度较DCNN网络提升了9.13%。而在进行肺结节早期预测时,准确率、召回率以及ZSI相比DCNN网络分别提升了0.052、0.039和0.043。

In order to improve the accuracy of early disease screening,this paper studies the computer aided diagnosis and treatment technology based on clinical test values,CT images and other data.Because medical images have the characteristics of serious artifacts and blurred edges,if the traditional Depth Learning(DL) network is used,the network depth needs to be increased,which puts forward higher requirements for computer computing power and the number of training samples.In order to overcome the above problems,this paper introduces the Reinforcement Learning(RL) algorithm,improves the two algorithms and proposes a DCQN network.The network combines and extracts the low order features of the image through the multi-layer nonlinear convolution network.With the help of the concept of Agent in RL,it continuously accumulates the penalty values in the process of environmental iteration,and finally obtains the optimal path to identify medical images.The simulation results on an open pulmonary nodule dataset show that the segmentation accuracy of DCQN network is 9.13% higher than that of DCNN network.In the early prediction of pulmonary nodules,the accuracy,recall and ZSI increased by 0.052,0.039 and 0.043 respectively compared with DCNN.