YU-net Lung Segment Image Preprocess Methods Used for Common Chest Diseases Prediction

Abstract

With the availability of large-scale data set of X-ray images and development of CNNs(Convolutional Neural Networks), using CNNs assist diagnose become more and more popular. But training CNNs using global image may be affected by the excessive irrelevant noisy areas. Due to the poor alignment of some Chest X-ray(CXR) images, the existence of irregular border hinders the neural network performance. In our work, we address the above problems by proposing a YU-net to segment lung fields on CXR images based on U-net, remove those areas of the images outside the lungs. In order to prove the effectiveness of YU-net, we trained, validated and tested the same 112,120 pictures of 30,536 patients on ResNet-50 and DenseNet-121 with both original Chest X-ray images and YU-net cleaned images. Compare the predicted result of DenseNet-121 and ResNet-50 with both YU-net processed images and original dataset, we found that use the YU-net cleaned images improve the performance of CNNs to recognize the multiple common thorax diseases.