Struski Łukasz, Urbańczyk Tomasz, Bucki Krzysztof, Cupiał Bartłomiej, Kaczyńska Aneta, Spurek Przemysław, Tabor Jacek
Faculty of Mathematics and Computer Science, Jagiellonian University, Kraków, Poland.
Skopia Medical Center, Kraków, Poland.
PLoS One. 2025 May 27;20(5):e0312038. doi: 10.1371/journal.pone.0312038. eCollection 2025.
The generation of videos is crucial, particularly in the medical field, where a significant amount of data is presented in this format. However, due to the extensive memory requirements, creating high-resolution videos poses a substantial challenge for generative models. In this paper, we introduce the Memory Efficient Video GAN (MeVGAN)-a Generative Adversarial Network (GAN) that incorporates a plugin-type architecture. This system utilizes a pre-trained 2D-image GAN, to which we attach a straightforward neural network designed to develop specific trajectories within the noise space. These trajectories, when processed through the GAN, produce realistic videos. We deploy MeVGAN specifically for creating colonoscopy videos, a critical procedure in the medical field, notably helpful for screening and treating colorectal cancer. We show that MeVGAN can produce good quality synthetic colonoscopy videos, which can be potentially used in virtual simulators.
视频生成至关重要,尤其是在医学领域,大量数据都是以这种格式呈现的。然而,由于对内存要求极高,生成高分辨率视频对生成模型来说是一项巨大挑战。在本文中,我们介绍了内存高效视频生成对抗网络(MeVGAN)——一种采用插件式架构的生成对抗网络(GAN)。该系统利用一个预训练的二维图像GAN,我们在其上附加了一个简单的神经网络,旨在在噪声空间中生成特定轨迹。这些轨迹经过GAN处理后可生成逼真的视频。我们专门将MeVGAN用于创建结肠镜检查视频,这是医学领域的一项关键程序,对结直肠癌的筛查和治疗非常有帮助。我们证明MeVGAN能够生成高质量的合成结肠镜检查视频,这些视频有可能用于虚拟模拟器。