Bridging the Density Gap: Diffusion Model for Stepwise Generation of Dense Cell Images from Sparse Data

Abstract

High annotation costs are a significant barrier in time-lapse cell imaging, as cell density increases over time, leading to numerous adjoining and difficult-to-distinguish boundaries. Our work is based on the observation that sparse images, which are less costly to annotate, still contain local regions with cell densities comparable to those in dense data. We aim to generate dense data from these sparse images for data augmentation. However, a simple generation approach fails due to the significant distribution gap between the training (sparse) and target (dense) data, resulting in unrealistic images. Therefore, we propose stepwise generation, a method that incrementally bridges this distribution gap. This approach successfully generates realistic images usable for augmentation, and our analysis confirms that the appearance of the generated cells are close to that of real data.

Kazuya Nishimura

Assistant Professor

Kazuya Nishimura is an assistant professor with D3 Center, the University of Osaka.