Publication

Abstract : Generating and detecting MRI images are most beneficial if the illness requires a fast and accurate cure. Concerning the weakness, there have been negative remarks made about DL for violating the privacy of patients. However, it takes a lot of money and time to collect a rich MRI image database to train the model. It has been found out that way too many medical imaging datasets contain imbalanced data, thus making it difficult for the model to find the outliers. Hence, the basic necessity in working with medical images is the process of data augmentation. The regular forms of data augmentation such as rotation, scale, crops, etc., results in images that look very similar and there isn’t a usual variation that is vital in DL algorithms to learn about the characteristics of the images. On the other hand, Generative Adversarial Networks (GAN) have reported the potentiality to generate synthetic data with good generalization to a large image set. It is also important to note that GANs also possess the property of being an cheap in terms of data processing. In this work, we adopted the AGG based on the characteristics of MRI data analysis, and PSNR for the analysis of local information of a source image based on style transfer and multiple GANs for shared information. Style transfer follows aggregation to ensure that the aggregated image is similar to the original images. Then, we perform an analysis of the aggregation and the style transfer and because it drastically reduced the performance, dropped them.