Generative Adversarial Networks for Anomaly Detection in Medical Images

Abstract

In computer vision, anomaly detection (AD) is a challenging task. AD presents additional difficulties, especially in the realm of medical imaging, for several reasons, one of which being the dearth of ground truth (annotated) data. AD models built on generative adversarial networks (GANs) have advanced significantly in the last several years. Their usefulness in biological imaging is still not well understood, though. In this study, we provide an overview of the use of GANs for AD and a detailed analysis of the difficulties faced in implementing the most advanced GAN-based AD techniques for biomedical imaging. Additionally, we have explicitly examined the benefits and constraints of AD approaches on medical image datasets, conducting tests on 2 medical imaging datasets from various modalities, organs, and tissues using 3 AD methods. We examined the outcomes from the perspectives of both data and models, given the strikingly disparate results obtained in these studies. The outcomes demonstrated that no technique could consistently identify anomalies in medical imaging. A few of the phenomena that have a significant influence on the AD models' performance include the quantity of training samples, the subtlety of the anomaly, and the anomaly's distribution throughout the images. We also anticipate significant research paths and offer recommendations for the application of AD models in medical imaging.