MicroGAN: Size-invariant Learning of GAN for Super-Resolution of Microscopic ImagesDownload PDF

Published: 08 Oct 2019, Last Modified: 05 May 2023COMPAY 2019Readers: Everyone
Keywords: Super-Resolution, Generative Adversarial Networks, Medical image analysis, Urine Sediment, Deep Learning
Abstract: One of the intrinsic problems in deep learning and related research in microscopic image analysis is the lack of availability of high quality images due to various factors such as limitations of optics, cost, etc. Further, there exists extreme variation in sizes of individual cells in microscopic data which leads to variability in the size of corresponding images. This demands for size independent inputs to the neural network which is in contrast to their traditional training procedures. In this paper, we propose MicroGAN, a generative adversarial network (GAN), for super resolving low quality images without ignoring the original size of individual cells, thus keeping intact their original texture and anatomy. Our method consists of a conditional GAN based on trimmed U-Net and WGAN-GP architecture followed by global average pooling for tackling the size-invariance problem during the learning phase. We tested our methodology on urine sediment analysis data which shows a significant perceptual as well as quantitative improvement. We also show an accuracy improvement of 2.37% in the intended classification task with the super resolved data, thus strongly verifying the proposed approach.
4 Replies

Loading