Cross-domain attention-guided generative data augmentation for medical image analysis with limited data

Zhenghua Xu, Jiaqi Tang, Chang Qi, Dan Yao, Caihua Liu, Yuefu Zhan, Thomas Lukasiewicz

Abstract: Highlights•Identify the problems in the traditional and existing GAN-based data augmentation works.•Propose a generative data augmentation solution, CDA-GAN.•Use a new AMSE block, a semi-supervised spatial attention, and spectral Normalization.•Propose a new strategy to extend the use of CDA-GAN to medical image segmentation tasks.•Experiments verify the effectiveness of our work in clinical medical image analysis tasks.