Feature Controlled Synthetic Medical Image Generation using Conditional Generative Adversarial Network

In the healthcare industry, the data available is very little, and due to privacy concerns, most Electronic Health Records (EHRs) are not shared. The synthetic generation of quality data from existing real medical data is done using generative models like Generative Adversarial Networks (GANs). The data scarcity and data privacy problems can be overcome through this synthetic data generation for medical data records. In this paper, the Conditional Generative Adversarial Network (CGAN) is CNN-based. It is a type of GAN where CGAN is used to generate realistic images by conditioning specific attributes such as age, gender, position and diseases to the generator and discriminator, unlike GAN, where the specific attributes (as labels) cannot be given along the input side. The Generator takes the controllable conditional attributes (age, gender, position, and diseases) as input alongside some noise and generates quality synthetic chest X-ray images, whereas the discriminator checks whether the condition matches the generated synthetic data with the real data. The model is trained with a large dataset with quality chest X-ray images. During the generation phase, the generation can be controlled by providing conditional attributes such as age, gender, position, and disease type. According to these given attributes, synthetic images are generated. Transfer learning (TL) [18] is used to train more epochs, as CGAN takes a longer training period. The CGAN model is evaluated in two ways: one is the qualitative analysis, in which visual inspection is done, and where in quantitative analysis, generated image quality is compared with real data quality using the metric Fréchet Inception Distance (FID).