Label-conditioned multi-GAN fusion: A robust data augmentation strategy for medical image segmentation

The performance of deep learning for medical image segmentation heavily relies on the quantity and quality of training data. However, lack of high-quality labeled data remains a critical bottleneck. It requires several hours of radiologist to annotate the organs in a CT/MRI. In addition, rare disease generally has limited samples for training, while anatomical boundary blurring and intra-class intensity heterogeneity also yield data scarcity on the other hand. To this end, this paper proposes a label-guided multi-GAN collaborative framework for medical image augmentation. Leveraging existing labels as conditional inputs, three GAN variants (Pix2pix, Pix2pixHD, SPADE) are trained in parallel to synthesize images in target domain. This design highlights anatomical regions, improves image quality, and enhances data diversity and quality at the same time. Experimental results on three modalities demonstrate that our approach is able to significantly boost segmentation performance across various segmentation networks.