Abstract
Denoising diffusion probabilistic models have shown great potential in generating realistic image data. We show how those models can be used to generate realistic microscopy image data in 2D and 3D based on simulated sketches of cellular structures. Multiple data sets are used as an inspiration to simulate sketches of different cellular structures, allowing to generate fully-annotated image data sets without requiring human interactions. Those data sets are used to train segmentation approaches and demonstrate that annotation-free segmentation of cellular structures in fluorescence microscopy image data can be achieved, thereby leaping towards the ultimate goal of eliminating the necessity of human annotation efforts.
Abstract (translated)
去噪扩散概率模型在生成真实图像数据方面展现出了巨大的潜力。我们展示了如何通过这些模型在2D和3D空间中基于模拟细胞结构 Sketch 生成真实显微镜图像数据。多个数据集用作灵感来模拟不同细胞结构的 Sketch,从而生成无需人类交互即可完全注释的图像数据集。这些数据集用于训练分割方法,并证明在荧光显微镜图像数据中无注释的细胞结构分割可以实现,从而朝着消除人类注释工作必要性的终极目标迈出了一大步。
URL
https://arxiv.org/abs/2301.10227
