融合位置尺度信息的胸部X光肺结节检测

Detection of Pulmonary Nodules on Chest X-ray with Fusion of Location and Scale Information

基于胸部正面X光的肺结节检测任务因结节较小、肋骨遮挡等原因检测难度较大,需要在保证高敏感度的前提下,尽可能地减少假阳性样本比率.目前大多数肺结节检测方法一般分为3个步骤:肺部区域分割;候选区域生成;通过进一步分类,减少假阳性结果.这类方法存在一些问题,每一步的结果都依赖于前一步的性能,整个流程往往会使用多个模型、多次处理以提升效果,算法复杂而且计算量大.同时,会有些结节因为器官遮挡不在肺部分割的区域内,肺部分割会漏掉一些结节.针对这个问题,本文使用一个端到端的目标检测网络来完成肺结节检测任务,X光片经过图像预处理后输入网络,直接得到肺结节的预测结果.此方法基于卷积神经网络(Convolutional Neural Network,CNN)的目标检测模型,同时在分类任务中融合位置和尺寸信息,实验证明这些信息有助于模型判断.在公开数据集--日本放射技术学会(Japanese Society of Radiological Technology,JSRT)数据集的实验结果显示,本文方法在平均每张图像4. 5个假阳性结果时敏感度为92%,2个假阳性结果时敏感度为88%,在较低的假阳性率的情况下,超出了先前的研究成果.

Pulmonary nodules on the chest X-ray film is difficult to be detected due to their small sizes and obstructions of ribs. Moreover,Detection methods need to ensure both high sensitivity and low false positive rate. Most pulmonary nodule detection methods based on the chest X-ray film generally involve three steps: lung region segmentation,candidate region generation and further classification to reduce false positive results. However,the results of each step depend on the performance of the previous step. Meanwhile,the whole process often uses multiple models to improve the effect,which is cumbersome and has a large amount of calculation. At the same time,due the organ occlusion,some nodules are not in the lung segmentation area. Thus lung segmentation preprocessing will miss some nodules. In this paper,an end-to-end neural network model is proposed. The network takes preprocessed X-ray film as the input and then get the prediction results of pulmonary nodules directly. This method is based on the object detection model of convolutional neural network( CNN),and adds location and size information to the classification task. Experiments show that it is useful.In the Japanese Society of Radiological Technology( JSRT) database,more than 88% and 92% of lung nodules in the lung field were detected when the false positives per image were 2 and 4. 5. In the case of a lower false positive rate,it exceeded the performance of the previous research results.