Functional-realistic CT image super-resolution for early-stage pulmonary nodule detection

Early-stage pulmonary nodule detection is challenging for Computer-aided Diagnosis systems (CADs) in clinical practice. It always relies on large-scale annotated pathological images. Unfortunately, the limited voxels of earlier-stage nodules can aggravate the risk of escaping diagnosis. Due to high-dose CT and bronchoscope potential threats, CT image super-resolution has become a suboptimal way to tackle the problem. Therefore, we proposed a deep generative adversarial network (GAN) architecture based on a deep grammar model, called FRGAN (Functional-Realistic GAN). By using region proposal network (RPN), the bottom semantic features are recommended and classified as the basic units of functional structure. Local pathological images can be hierarchically aggregated with corresponding to different semantic patterns as parse trees. Refer to their EMRs, and we use TreeGAN to generate the correct syntax patterns for each early-stage pulmonary nodule candidates. We report the generated results of the super-resolution images and feed them into a convolutional network to assess the functional loss of the generated results along to different parse trees. Within the contextual and generative losses, we rebuild a novel objective function paralleling with TreeGAN. The aim is to boost the sensibility of pulmonary nodule detection with more functional-realistic data augmentation. Experimental results show that our generator can faster generate more realistic SR images with pathological features. Moreover, it could be a data augmentation tool for some deep architecture to overcome sample imbalance.