Use Data Augmentation Model in Image Segmentation

In our research, we have harnessed the power of deep learning techniques for data augmentation. Specifically, we employed generative adversarial network (GAN) models to perform image generation tasks. The Pix2Pix model is an improved version of GAN that we have proposed due to its outstanding ability in generating images effectively. The Pix2Pix model is based on the U-Net architecture, a well-known network architecture in biomedical image segmentation. The layers in the left branch (encoder part) will be concatenate directly into the layers in the right branch (decoder part) of the same size, and this connection is called skip connection to add more information and increase accuracy. During the experimental study, we found that the Pix2Pix model is still able to do better by adding Attention Gate to this model because the model trained with Attention Gates completely learns to remove the Attention Gates as well as unrelated regions in the input image while highlighting salient features which are useful for a specific task. Moreover, the Attention Gate can be easily integrated into standard CNN architectures such as the U-Net model. For that reason, we have proposed a new approach by adding Attention Gate at the skip connections in the generator model of Pix2Pix to augmentation of the lung’s CT image data for our Covid-19 disease classification problem. In the paper, we used an initial training dataset consisting of 17,000 lung CT image files and 16,700 lung CT mask image files. The results indicate that the original Pix2Pix model achieves a quality score (FID) of 83.94343628014835. The data after being enhanced by the Pix2Pix model is of sufficient quality to be used in the identification of Covid-19. Concurrently, we also conducted experiments with the Pix2Pix Attention U-Net generator model and obtained promising results, reaching a quality score (FID) of 103.15617916362004. The results show that the combination of Attention Gate into the U-Net generator of the Pix2Pix model has generated reliable lung CT image data.