Task Driven Generative Modeling for Unsupervised Domain Adaptation: Application to X-ray Image Segmentation

Motivation

• if we don’t have any label in the target domain, how can we train a model to do segmentation or other tasks?
• we have lots of one modality (CT) paired (labeled) data in hands.  If it is possible to train a model with unsupervised learning to segment the orans from x-ray (target) images?

Solution

• train a segmentation model on CT labeled data (supervised learning)—DI2I
• use a generative model to synthesize fake-CT from X-ray (fake and real)
• then send the face-CT to DI2I for segmentation
• In short:  X-ray ->Generator->  Fake CT –>DI2I
• ->Cycle-GAN for synthetic 
• ->concatenate the loss from DI2I with the loss from discriminator D2

 

 

 

 

paper: https://arxiv.org/abs/1806.07201