A capstone project dedicated to leveraging Style GAN (Generative Adversarial Network) to generate Brain MRI images of different contrasts
To build a Generative adversarial model(modified U-Net) which can generate artificial MRI images of different contrast levels from existing MRI scans.
1. Discriminator : A classifier network that identfies whether a generated data (say, image) is fake or not.
2. Generator : A network that generates fake data (say, image) from random noise samples. It's objective is to generate surreal data and fool the discriminator network.
1. Discriminator Loss = The loss associated with classifying the real-image and the generated fake image.
2. Generator Loss = The loss associated with generating the fake image, to make it look surreal, .i.e. exclusively on the generator output.
3. Objective Function
-> Here, we have to maximize the likelihood for discriminator being wrong
-> The discriminator tries to maximize the objective function (Gradient Ascent)
-> The generator tries to minimize the objective function (Gradient Descent)