UDCT: Unsupervised data to content transformation with histogram-matching cycle-consistent generative adversarial networks

Abstract

The segmentation of images is a common task in a broad range of research fields. To tackle increasingly complex images, artificial intelligence (AI) based approaches have emerged to overcome the shortcomings of traditional feature detection methods. Owing to the fact that most AI research is made publicly accessible and programming the required algorithms is now possible in many popular languages, the use of such approaches is becoming widespread. However, these methods often require data labeled by the researcher to provide a training target for the algorithms to converge to the desired result. This labeling is a limiting factor in many cases and can become prohibitively time consuming. Inspired by Cycle-consistent Generative Adversarial Networks' (cycleGAN) ability to perform style transfer, we outline a method whereby a computer generated set of images is used to segment the true images. We benchmark our unsupervised approach against a state-of-the-art supervised cell-counting network on the VGG Cells dataset and show that it is not only competitive but can also precisely locate individual cells. We demonstrate the power of this method by segmenting bright-field images of cell cultures, a live-dead assay of C.Elegans and X-ray-computed tomography of metallic nanowire meshes.

The role of convolutional neural networks in scanning probe microscopy: a review

“…With training, the discriminator improves, and the generator learns to generate fake images that look real. This process repeats itself until the discriminator cannot distinguish between the fake images generated and the real ones [103,104]. GANs have been used to achieve segmentation without human annotation [104].…”

“…This process repeats itself until the discriminator cannot distinguish between the fake images generated and the real ones [103,104]. GANs have been used to achieve segmentation without human annotation [104]. In this approach, computer-generated images are used for segmenting true bright-field images by synthesizing colorful segmented images and then overlaying them on the experimental image.…”

GAN based Unsupervised Segmentation: Should We Match the Exact Number of Objects

“…However, the quality of synthesis between real images and clean binary masks is typically not optimal since the underlying Poisson distribution of the binary masks is not a realistic distribution in real images. 13 Moreover, the optimization of the KL divergence for training discriminators is more difficult to converge 14 using clean binary masks. Therefore, the Gaussian smoothing, random noise, and brightness variations are used to generate augmented masks M A in additional to the simulated mask images M for better synthetic performance (Fig.…”

“…[9][10][11][12] Within the CycleGAN framework, 7 many previous studies have tackled the unsupervised semantic segmentation in microscopy imaging. Ihle et al 13 proposed to use the CycleGAN framework to segment bright-field images of cell cultures, a live-dead assay of C.Elegans, and X-ray-computed tomography of metallic nanowire meshes. A similar approach was proposed by 14 for facilitating stain-independent supervised and unsupervised segmentation on kidney histology.…”

“…However, the previous studies emphasized Further author information: Quan Liu : E-mail: quan.liu@vanderbilt.edu Contact Author: Yuankai Huo: E-mail: yuankai.huo@vanderbilt.edu that the macro-level (global distribution level) matching on the number of objects between intensity images and simulated masks improved the segmentation performance. 13 That fact inspired us with the question that if the segmentation performance could be further improved by doing more careful matching than the macro-level. To answer the question, we propose a new micro-level matching (mini-batch level) strategy to match the rough number of objects across two domains when training the CycleGAN framework.…”