融合迭代自组织数据分析的肺结节自动检测算法

An Automatic Pulmonary Nodule Detection Algorithm Based on Iterative Self-organizing Data Analysis

肺结节的早期诊断能为肺癌的及时发现提供可能,而肺部断层扫描(Computed Tomography,CT)为结节的诊断提供了一种便利方法。为了辅助医生对CT中肺结节的有效读取,论文提出了一种基于卷积神经网络(Convolutional Neural Network,CNN)的肺结节自动检测算法。该算法以Faster R-CNN网络为基础框架,首先,使用迭代自组织数据分析算法(Iter-ative Self-organizing Data Analysis Techniques Algorithm,ISODATA)对所输入的肺部CT图像进行聚类学习,以自动生成更加符合肺结节形态特征的锚框参数,进而指导区域候选网络生成更加贴合结节真实大小的候选框,在减少假阳结果产生的同时,提高肺结节检测的精确度;其次,采用Focal Loss作为分类损失函数,其通过对难以识别的肺结节样本分配更高的权重,以加强对这些样本的学习,进而提高算法的分类能力。与目前先进的目标检测算法相比较,论文所提出的算法在肺结节数据集LIDC-IDRI上,取得了优异的表现,并且在对正确分辨肺结节的真伪性方面具有更大的潜力。

Early diagnosis of pulmonary nodules can provide possibility for early detection of lung cancer.Computed tomogra-phy(CT)provides a convenient method for the diagnosis of pulmonary nodules.In order to assist doctors to read lung nodules in CT,an automatic detection algorithm based on convolutional neural network(CNN)is proposed in this paper.Based on the Faster R-CNN method,an iterative self-organizing data analysis techniques algorithm(ISODATA)is firstly presented.ISODATA,based on the clustering learning of the input lung CT images,automatically generates anchor frame parameters that are more in line with the morphological characteristics of pulmonary nodules and then guides the regional candidate network to generate candidate frames that are much closer to the real size of the nodules,so as to reduce the false positive results while improving the accuracy of pulmo-nary nodules detection.Secondly,Focal Loss is used for the classification,and relatively high weights are assigned to the lung nodu-lar samples that are difficult to identify so as to enhance the learning of these samples,and further improve the classification ability of the algorithm.The experimental results have demonstrated that the proposed algorithm on the open LIDC-IDRI dataset is superior to the state-of-the-art algorithms.It implies great potential in correctly distinguishing the authenticity of pulmonary nodules.