Image data augmentation for improving performance of deep learning-based model in pathological lung segmentation

ABSTRACT

Accurate segmentation of lung fields from chest Xray (CXR) images is very important for subsequent analysis of many pulmonary diseases. Deep Neural Networks (DNN)-based methods have achieved remarkable progress in many image related tasks. However, their performance depends highly on the distribution of training and test samples, and they perform well if both training and test samples are from the same distribution. For example, DNN-based lung segmentation methods perform well on segmentation of healthy lung or lung with mild disease, however their performance is poor on lungs with severe abnormalities. Pulmonary opacification, which blurs the lung boundary, is one of the main reasons. A solution to this problem is data augmentation to increase the pool of training images, however despite the great success of traditional data augmentation techniques for natural images, they are not very effective for medical images. To simulate CXR images with opacification and low contrast, we present a novel image data augmentation technique in this study. To generate an augmented image, we first generate a random area inside the lung and then blur the area with a gaussian filter. Then, low contrast is simulated by adjusting the contrast and brightness. To evaluate the utility of the proposed augmentation technique, we applied it to images with different pulmonary diseases such as tuberculosis, pneumoconiosis and covid-19 from three public datasets as well as a private dataset and compared its effect on segmentation performance with traditional data augmentation techniques. Results suggest that the proposed technique outperforms traditional data augmentation techniques for all datasets on lung segmentation, in terms of Dice Coefficient (DC) and Jaccard Index (JI). Extensive experiments on multiple datasets validate the effectiveness of the proposed data augmentation technique.