RA-GAN data enhancement method to improve the performance of lung parenchyma segmentation

Poor generalization and low accuracy of the neural network for segmentation is often lead by a small medical images dataset during model training for segmentation tasks. That is the mainly reason that data enhancement methods are usually needed to augment the dataset. Currently, data enhancement methods based on generative adversarial networks generate unstable images with local blurring and unclear edge distinction, which lead to the insignificant improvement of segmentation performance after data enhanced medical images. The data enhancement method is proposed which based on improved conditional generative adversarial network for augmenting lung CT images in this paper, in order to effectively solve this problem. The network depth is deepened by introducing residual blocks, and the CBAM attention module is added in the jump connection stage to obtain richer global features; the ViT-based (Vision Transformer) discriminator is used to guide the generator to generate CT images with higher quality and more distinct edge differentiation. A segmentation experiments are conducted by data enhancement to verify the effectiveness of the method. The experimental results show that the method performs better in the small sample segmentation task compared to other segmentation methods.