Chest X-ray Abnormality Detection Using Convolutional AutoEncoder Combined with Double Generative Adversarial Network (GAN)

The statistical properties of aberrant samples are unstable, and traditional chest X-ray image data is difficult to gather and unevenly distributed. A CAE-D-GAN anomaly detection model based on dual GANs and convolutional autoencoders is proposed in this paper. We employ the DGAN module to obtain clear degraded images, the convolutional autoencoder to extract low-dimensional features from high-dimensional data, and the DDGAN module to learn how to degrade images and recover clearer images from degraded photos during the adversarial process. While the reconstruction error score identifies sample flaws, the network is optimized by the error loss between the original and reconstructed samples. The network just needs normal samples to be trained, and it can reach a maximum AUROC value of 0.86. The findings demonstrate that the CAE-D-GAN model outperforms a number of different anomaly detection models in terms of detection effects and feature reconstruction capabilities. There are special opportunities for using this approach to detect other anomalies in medical images.