TopoSegNet: Scalable topology preservation in image segmentation via critical points.
Saved in:
| Title: | TopoSegNet: Scalable topology preservation in image segmentation via critical points. |
|---|---|
| Authors: | Ahmadkhani, Mohsen1 (AUTHOR) ahmad178@umn.edu, Shook, Eric1 (AUTHOR) eshook@umn.edu |
| Source: | Computer Vision & Image Understanding. Dec2025, Vol. 262, pN.PAG-N.PAG. 1p. |
| Subjects: | Image segmentation, Critical point theory, Deep learning, Optimization algorithms |
| Abstract: | Image segmentation is crucial in computer vision, with applications in various fields. Despite advancements in deep learning techniques, maintaining topological consistency in segmented outputs remains a significant challenge. Traditional topology-aware methods, such as those using persistent homology (PH), preserve topological features such as loops and connected components but are often computationally expensive. We introduce a new approach that shifts focus from topological features, such as loops and connected components, to points. We term points that play a crucial role in topology, topologically critical points (TCPs), such as thin segments, junctions, and terminal nodes that form these topological features. Shifting attention to points reduces computational overhead while capturing the most critical topological structures. We propose TopoSegNet, a novel topology-aware loss function for image segmentation that emphasizes the preservation of TCPs to maintain topological integrity without the computational complexity of traditional approaches. We also propose the Centroid Displacement Measure (CDM), a new evaluation metric to quantify topological and geometric fidelity. Across multiple datasets, TopoSegNet demonstrates consistent performance gains compared to the three baseline models. On GMP (tree rings), it improves dice, CDM, and MMD by 4%, 17%, and 19%, respectively, while reducing execution time from over 21 h to 3.4 h. On DRIVE (retinal vessels), it achieves gains of 7%, 19%, and 16% on the same metrics. On CREMI-B (neuronal structures), TopoSegNet improves CDM by 14.77% and dice by 3.3%, while on the satellite imagery of agricultural fields, it achieves a 24.48% CDM improvement with a slightly higher dice score (42.8 vs. 42.74) compared to the baseline topology-aware model. These results highlight TopoSegNet as an efficient and scalable solution for topology-aware image segmentation. • Introduced TopoSegNet for topology-aware image segmentation using topologically critical points (TCPs). • Proposed the application of 0-simplices (TCP) as a scalable alternative for persistent homology (PH). • Improved topological and pixel-wise accuracies in the image segmentation task. • Proposed Centroid Displacement Measure (CDM) for topological evaluation. [ABSTRACT FROM AUTHOR] |
| Copyright of Computer Vision & Image Understanding is the property of Academic Press Inc. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Engineering Source |
| Abstract: | Image segmentation is crucial in computer vision, with applications in various fields. Despite advancements in deep learning techniques, maintaining topological consistency in segmented outputs remains a significant challenge. Traditional topology-aware methods, such as those using persistent homology (PH), preserve topological features such as loops and connected components but are often computationally expensive. We introduce a new approach that shifts focus from topological features, such as loops and connected components, to points. We term points that play a crucial role in topology, topologically critical points (TCPs), such as thin segments, junctions, and terminal nodes that form these topological features. Shifting attention to points reduces computational overhead while capturing the most critical topological structures. We propose TopoSegNet, a novel topology-aware loss function for image segmentation that emphasizes the preservation of TCPs to maintain topological integrity without the computational complexity of traditional approaches. We also propose the Centroid Displacement Measure (CDM), a new evaluation metric to quantify topological and geometric fidelity. Across multiple datasets, TopoSegNet demonstrates consistent performance gains compared to the three baseline models. On GMP (tree rings), it improves dice, CDM, and MMD by 4%, 17%, and 19%, respectively, while reducing execution time from over 21 h to 3.4 h. On DRIVE (retinal vessels), it achieves gains of 7%, 19%, and 16% on the same metrics. On CREMI-B (neuronal structures), TopoSegNet improves CDM by 14.77% and dice by 3.3%, while on the satellite imagery of agricultural fields, it achieves a 24.48% CDM improvement with a slightly higher dice score (42.8 vs. 42.74) compared to the baseline topology-aware model. These results highlight TopoSegNet as an efficient and scalable solution for topology-aware image segmentation. • Introduced TopoSegNet for topology-aware image segmentation using topologically critical points (TCPs). • Proposed the application of 0-simplices (TCP) as a scalable alternative for persistent homology (PH). • Improved topological and pixel-wise accuracies in the image segmentation task. • Proposed Centroid Displacement Measure (CDM) for topological evaluation. [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 10773142 |
| DOI: | 10.1016/j.cviu.2025.104564 |
