2026-03-27 東京大学,京都大学,科学技術振興機構
トポロジカルデータ解析に基づく細胞膜画像セグメンテーションツール「PomSeg」 （論文グラフィカルアブストラクトより）
＜関連情報＞
- https://www.jst.go.jp/pr/announce/20260327/index.html
- https://www.jst.go.jp/pr/announce/20260327/pdf/20260327.pdf
- https://www.cell.com/cell-reports-methods/fulltext/S2667-2375(26)00066-4
膜画像のための持続的な相同性に基づくセグメンテーションツール Persistent homology-based segmentation tool for membrane images
Haruhisa Oda ∙ Yusuke Imoto
Cell Reports Methods Published:March 26, 2026
DOI:https://doi.org/10.1016/j.crmeth.2026.101366
Motivation
The trade-off between flexibility and automation is a crucial topic in cell segmentation. Machine learning-based tools serve as examples of automation technologies, designed to learn from and adapt to input datasets. However, there is also a significant need for user-centered, flexible tools that have parameters with clear meanings. Persistent homology is an advanced method used to capture topological information from data, and it has been employed to create user-centered cell nuclei segmentation tools. Despite this, there has yet to be a development of persistent homology-based membrane image segmentation tools with user-centered features, even though this image modality is vital for biological studies. Here, we introduce PomSeg, a persistent homology-based membrane image segmentation tool that utilizes both 2D and 3D persistent homology processes.
Highlights
- PomSeg, a persistent homology-based membrane image segmentation tool
- User-centered flexibility and versatility ensures control in image analysis tasks
- High consistency with ground truth and robustness to noise and resolution
- PomSeg offers precision and control for image analysis users
Summary
While advances in machine learning have enabled automated cell segmentation, users often face challenges in parameter tuning until reaching their desired results. To address this issue, we developed PomSeg, a membrane segmentation method based on persistent homology. Since persistent homology captures topological features of input data, PomSeg parameters reflect cell shape information, enabling intuitive and efficient parameter tuning. This adaptivity, together with stability for noise of persistent homology, enables robust application of PomSeg to various image types, including coarse-resolution data. By applying PomSeg to early mouse embryo membrane images and other publicly available datasets, we demonstrated its flexibility, versatility, and robustness, along with agreement with ground truth. Additionally, we showed the potential of PomSeg extension by incorporating a machine learning tool in its process. These features make PomSeg a valuable tool for researchers pursuing control and interpretability in segmentation, as well as indicating wider applications beyond a machine learning alternative.
