Integrating SSIM in GANs to Generate High-Quality Brain MRI Images

Abstract

Application of machine learning (ML) techniques in medical imaging is not new, but modern techniques like deep learning require a lot of quality data to perform effectively. Data, present in the field of medical imaging, imposes two main issues: first is the lack of anomalous data. For instance, only a few medical resonance images of the brain with tumor are present as compared to that of a healthy human brain. This makes the task of anomaly detection using ML techniques suffer from a lack of data. Second, the data that is present is highly confidential, and most of the time, the patient does not want it published anywhere. Many data augmentation techniques have been proposed in the past, but such techniques do not create data that is completely different from the given original data. In our work, we focus on using a novel extension of deep convolutional generative adversarial networks (DCGANs) to generate high-resolution brain magnetic resonance (MR) images of both anomalous and normal brain. This task is a tricky one as a tumor is present at random spots of the brain and has random sizes. For this reason, vanilla generative adversarial networks (GANs) cannot generate realistic tumor images. We propose a novel approach that integrates structural similarity index (SSIM) score in the objective function which, as we shall see, does a much better job at generating anomalous brain MR images as compared to vanilla GANs.