基于卷积神经网络的肺野分割和肺炎筛查

Lung field segmentation and pneumonia screening using convolutional neural networks

由于X光胸片影像受到各类组织阴影及病灶的影响,目前肺野分割算法的结果往往存在空洞或者边缘不光滑等问题。针对此类问题,该文提出了一种基于多尺度卷积和特征金字塔的肺野分割网络,此网络利用多尺度卷积模块和多尺度特征融合模块提取和融合多尺度特征,在JSRT数据集上PA和Dice指标分别达到98.76%和97.94%,在Montgomery数据集上PA和Dice指标分别达到了98.96%和97.85%。该文将肺野分割网络进一步应用到肺炎筛查任务中,提出了一种基于肺野分割的数据增强方法,该方法通过分别“擦除”左右肺部的随机区域增加了样本的多样性,从而提高了肺炎分类任务的准确率。实验表明,这种数据增强方法可以将新冠肺炎检出率至少提高2.2%。

Because chest X-rays are affected by various tissue shadows and lesions,the results of current lung field segmentation algorithms often suffer from problems such as holes or unsmooth edges.To solve this problem,a lung field segmentation network based on multi-scale convolution and feature pyramid is proposed in this paper.This network uses multi-scale convolution module and multi-scale feature fusion module to extract and fuse multi-scale features.The Dice values of this network on JSRT and Montgomery datasets are 97.94%and 97.85%respectively and the pixel accuracy of this network on JSRT and Montgomery datasets are 98.76%and 98.96%respectively.Furthermore,we apply the lung field segmentation network to the pneumonia classification task,and propose a data augmentation method based on lung field segmentation.This data augmentation method increases the diversity of samples by erasing the random regions of the left and right lungs,thereby the accuracy of the pneumonia classification task is improved.Experiments show that this data augmentation method can increase the COVID-19 detection rates by at least 2.2%.