Generative Adversarial Framework for Learning Multiple Clinical Tasks

Inspired by the recent success of generative adversarial networks (GANs), we propose a novel adversarial network for learning multiple clinical tasks. We design a conditional generative adversarial network (cGAN) with a new selective weighted loss to mitigate imbalance data problem in medical image analysis. Our proposed method comprises two components: a generator and a discriminator. While the generator is trained on sequential magnetic resonance images (MRI) to learn semantic segmentation and disease classification, the discriminator classifies whether a generated output is real or fake. The generative model and the discriminator model are trained via adversarial loss with two player mini-max game, and with an additional proposed selective weighted loss. The proposed architecture has shown promising performance on the ACDC-2017 benchmark for prediction of cardiac disease beside of semantic segmentation of dual cavities and myocardium vessel. Moreover, we achieved competitive results for brain tumor semantic segmentation and brain disease classification on the BraTS-2017 challenge.