Tracking Temporal Evolution of Topological Features in Image Data
Susan Glenn, Jessi Cisewski-Kehe, Jun Zhu, William M. Bement · 2025
Read Tracking Temporal Evolution of Topological Features in Image Data by Susan Glenn, Jessi Cisewski-Kehe, Jun Zhu, William M. Bement (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-17530.txt.
Introduction
Susan Glenn, Jessi Cisewski-Kehe, Jun Zhu, and William Bement present a statistical method for image series, in two thousand twenty-five. Images can hold holes and loops, and tools exist to find such shapes in a single frame. What was missing was a sound way to track those shapes as they change over time.
Their framework tests for real shapes across both space and time at once. It links shapes in single frames into higher forms that show change over time. A diagram then sums up how each shape emerges and moves on. The method captured a wounded cell as it healed, and it beat a current approach in simulated tests.
Catalog record
| CREATOR | Susan Glenn, Jessi Cisewski-Kehe, Jun Zhu, William M. Bement |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 68 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-17530
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Tracking Temporal Evolution of Topological Features in Image Data Susan Glenn, Jessi Cisewski-Kehe, Jun Zhu, William M. Bement Abstract Topological Data Analysis (TDA) can be used to detect and characterize holes in an image, such as zero-dimensional holes (connected components) or one-dimensional holes (loops). However, there is currently no widely accepted statistical framework for modeling spatiotemporal dependence in the evolution of topological features, such as holes, within a time series of images. We propose a hypothesis testing framework to identify statistically significant topological features of images in space and time, simultaneously. This addition of time may induce higher-dimensional topological features which can be used to establish temporal connections between the lower-dimensional features at each point in time. The temporal evolution of these lower-dimensional features is then represented on a zigzag persistence diagram, as a topological summary statistic focused on time dynamics. We demonstrate that the method effectively captures the emergence and progression of topological features in a study of a series of images of a wounded cell as it repairs. The proposed method outperforms a current approach in a simulation study that includes features of the wound healing process. Since, the wounded cell images exhibit nonlinear, dynamic, spatial, and temporal structures during single-cell repair, they provide a good application for this method. keywords [class=MSC] Primary 62R40 ; secondary 62M10 , 62M40 keywords Image Processing , Time Series , Topological Data Analysis , Zigzag Persistence , and ?? Los Alamos National Laboratory, Los Alamos, NM, USA, ?? ?? Department of Statistics, University of Wisconsin, USA, ?? ; ?? ?? Center for Quantitative Cell Imaging, Department of Integrative Biology, University of Wisconsin, USA , ?? 1 Introduction Topological data analysis (TDA) provides a flexible framework for capturing nonlinear structure in image data by characterizing shape through connectivity and holes. A hole in an image is a topological feature characterized by a specific dimension, such as, a connected component (dimension 0), a loop (dimension 1), or a void (dimension 2), etc. The overall hole structure, which reflects holes across different dimensions, can be summarized using topological summary statistics. There have been many successful applications of TDA to understanding complex patterns in images, from neuroscience to climate science, and much more [ 31 , 39 , 40 , 41 , 12 , 42 ] . Extending these ideas to temporal data typically involves either computing time-varying topological summary statistics or applying TDA to lower-dimensional embeddings that capture the underlying dynamics. Images, in general, are challenging as they are typically multidimensional with non-linear patterns. Spatiotemporal image data add further complexity due to the combination of both spatial dependencies (e.g., nearby pixels in an image are more related than pixels further away), temporal dependencies (e.g., recent images are more related to present images than those in the more distant past), and their interactions. Despite advances in statistical techniques to disentangle these multi-scale dependencies in non-linear, high-dimensional spaces [ 11 ] , it remains a challenge to detect and characterize topological features of images evolving over time. Techniques in TDA for tracking temporal dynamics in data include sliding window embeddings, zigzag persistence, crocker plots, multiparameter persistence, and vineyards. However, most were developed for data such as one-dimensional times series, networks, point clouds, or binary images, and do not directly generalize to sequences of grayscale images [ 15 , 17 , 18 , 19 , 20 , 22 , 23 , 38 ] . A sliding window embedding maps overlapping time domain windows of a time series to a point-cloud representation and is most commonly used with scalar time series. While some work has extended these techniques to time series of higher-dimensional data structures, such as images or networks, they usually either (i) reduce the images or graphs to a scalar-valued function and then track temporal changes, or (ii) capture only global temporal dynamics throughout a video as opposed to specific features [ 17 , 35 , 36 , 37 , 44 ] . Similarly, zigzag persistence provides a flexible framework to track dynamic topological features across time but cannot be directly applied to sequences of grayscale images without preprocessing steps such as binarization or conversion to point clouds to extract meaningful topological structure [ 19 , 13 ] . Crocker plots, while capable of tracking the number of topological features over time, lack the ability to distinguish individual feature trajectories [ 22 ] . Another possible option is multiparameter persistence which allows for tracking changes in topological features across multiple scales and dimensions, such as spatial and temporal scales [ 34 ] . An extension of multiparameter persistence to a time series of images could potentially identify changes to a hole structure over time; however, this approach remains under development and poses significant computational and theoretical challenges. The field is still working toward computationally efficient, robust, and widely applicable techniques that can handle complex data, such as time series of images. The challenges described above in analyzing a general time series of images using TDA can be seen in our data application: tracking structural changes to a cell’s wound over time. Wound healing is a dynamic process by which a protein ring forms around the wound site and gradually shrinks as the cell heals. Several topological techniques exist for tracking the feature that represents the wound over time, and the feature’s persistence can be used to quantify the ‘size’ or ‘prominence’ of that feature. In [ 12 ] , a method was developed to estimate the persistence and uncertainty of the cell wound in an image at a single time point, and by using domain knowledge, these topological features (loops) could be connected across time, assuming the continuity of the wound. In contrast, [ 13 ] and [ 38 ] explicitly incorporate temporal dynamics by tracking specific topological features over time through their persistence. Each of these techniques provide information limited to changes in the persistence of individual loops over time, under the assumption that loops remain non-overlapping in their persistence values. However, these methods do not offer information about how loops interact over time, such as splitting, merging, or intersecting. Building on existing methodology like zigzag persistence, we propose a new approach that captures changes in the wound topology as it evolves. To address this goal of tracking the temporal evolution of a topological feature such as a cell wound, we propose a new approach for analyzing a time series of grayscale images as a three-dimensional dataset (i.e., the two-dimensional image, time, and the pixel intensity value in the spatiotemporal grid). We treat time as an intrinsic dimension of the data (rather than a dimension of the topological summaries of the data), allowing us to identify high-dimensional topological features (e.g., loops connected through time) directly from the image stack. We introduce a novel hypothesis testing framework to determine which of these features are statistically significant and to determine a threshold for the data (rather than using a predetermined threshold). This approach produces a binary image stack that preserves meaningful spatial and temporal structure and is grounded in statistical theory. Tracking multiple topological features across time is important in the context of TDA because it addresses a key limitation in many existing approaches: loss of a unique