How much BiGAN and CycleGAN-learned hidden features are effective for COVID-19 detection from CT images? A comparative study

Abstract

Bidirectional generative adversarial networks (BiGANs) and cycle generative adversarial networks (CycleGANs) are two emerging machine learning models that, up to now, have been used as generative models, i.e., to generate output data sampled from a target probability distribution. However, these models are also equipped with encoding modules, which, after weakly supervised training, could be, in principle, exploited for the extraction of hidden features from the input data. At the present time, how these extracted features could be effectively exploited for classification tasks is still an unexplored field. Hence, motivated by this consideration, in this paper, we develop and numerically test the performance of a novel inference engine that relies on the exploitation of BiGAN and CycleGAN-learned hidden features for the detection of COVID-19 disease from other lung diseases in computer tomography (CT) scans. In this respect, the main contributions of the paper are twofold. First, we develop a kernel density estimation (KDE)-based inference method, which, in the training phase, leverages the hidden features extracted by BiGANs and CycleGANs for estimating the (a priori unknown) probability density function (PDF) of the CT scans of COVID-19 patients and, then, in the inference phase, uses it as a target COVID-PDF for the detection of COVID diseases. As a second major contribution, we numerically evaluate and compare the classification accuracies of the implemented BiGAN and CycleGAN models against the ones of some state-of-the-art methods, which rely on the unsupervised training of convolutional autoencoders (CAEs) for attaining feature extraction. The performance comparisons are carried out by considering a spectrum of different training loss functions and distance metrics. The obtained classification accuracies of the proposed CycleGAN-based (resp., BiGAN-based) models outperform the corresponding ones of the considered benchmark CAE-based models of about 16% (resp., 14%).

A Comparative Study of Generative Adversarial Networks for Generating Car Damaged Images

“…In consideration of the network capacity to defeat the aforementioned cases, Tang, H. [36] propose unsupervised AttentionGAN with attention guided image-to-image translation to identify the foreground of the target domain and minimize the background changes of the source domain with two proposed AGs and ADs by generating foreground attention masks and a background attention mask and achieving good results over GANimorph [37] and CycleGAN. Now the researchers have been interesting in not only a comparative study of different GANs with metric evaluation methods but also a combination of them with CNN classification models to classify the specific tasks, especially in car generation and classification, Faster RCNN [58] and day to night transfer based on unsupervised CycleGAN were adopted as a cross domain car detection and generation in [38], in medical image synthesis: Sarv Ahrabi, S. in 2022 [39] presented the better performance results of CycleGAN by a comparative study among CycleGAN and BiGANs [59], as similarly outperforming results of CycleGAN as a comparative analysis among CycleGAN and GANs [60] in [40], and five different GANs such as CGAN [61], DCGAN [62], f-GAN [63], WGAN [64]; CycleGAN by comparison in [41], using CT images for COVID-19 detection with their classification accuracies and FID values, and in face-aging application, Sharma, N. in 2022 [42] observed that the overall performance of CycleGAN was better than AttentionGAN for face-recognizing with age progression by analyzing CycleGAN and AttentionGAN with CelebA-HQ (CelebFaces Attributes high-quality dataset) and FFHQ (Flickr Faces HQ). Nevertheless, comparing GAN performances with the evaluation ways of their generated images for many specific tasks in different application areas has remained on the stage of the unfinished debate.…”

“…Now many numerous architectures of GANs are available to perform and the researchers have been passionate about not only a comparative study of different GANs models with metric evaluation methods but also their combinations with CNN classification models to classify the specific tasks in different application areas. As far as we know outperforms CycleGAN over other deep generative models by a comparative analysis in car domain adaption from day to night translating [38], medical image synthesis [39][40][41], and face-aging application [42]. Nevertheless, comparing GAN performances with the evaluation ways of their generated images for many specific tasks in different application areas has remained on the stage of the unfinished debate.…”