Nodule Synthesis and Selection for Augmenting Chest X-ray Nodule Detection

Abstract

Nodule detection in chest X-ray (CXR) images is important for early screening of lung cancer. It typically requires a large number of well-annotated data to train an effective nodule detector. However, high-quality annotations are hard to obtain due to the difficulty of locating nodules in CXR images and high cost of recruiting experienced radiologists. To address this issue, we propose an inpainting-based data augmentation (DA) framework, which consists of Nodule Synthesis stage and Nodule Selection stage, to synthesize CXR images with plausible nodules for facilitating the subsequent task of nodule detection. A partial convolutional U-Net is applied in Nodule Synthesis stage, which can offer flexibility to generate nodules at various locations in lungs. Since not all the synthesized CXR images are effective for data augmentation, we introduce Nodule Selection stage to identify efficacious nodules from the synthesized CXR images, to effectively augment the variety of training data for nodule detection. Our experimental results show that our DA framework can produce synthesized CXR images with plausible nodules of high quality, whereas the data augmentation can significantly improve the nodule detection performance.