Dual-Topological Learning for Structural Integrity Preservation in Image Segmentation
| Reference No. | 2026a035 |
|---|---|
| Type/Category | Grant for General Research-Short-term Joint Research |
| Title of Research Project | Dual-Topological Learning for Structural Integrity Preservation in Image Segmentation |
| Principal Investigator | KANG MEIYAN（Kyoto University・Post Doctor） |
| Research Period |
July 27,2026. -
July 31,2027. |
| Keyword(s) of Research Fields | Image Segmentation, Topological Learning, Persistent Homology, Deep Learning, Medical Imaging, GPR, Structural Preservation |
| Abstract for Research Report |
This research aims to develop a novel image segmentation framework that preserves both topological consistency and geometric fidelity. In many real-world applications, such as medical imaging and subsurface structure analysis using ground-penetrating radar (GPR), maintaining structural integrity during segmentation is essential. However, conventional deep learning-based segmentation methods often fail to preserve thin structures and connectivity, leading to unreliable results. To address this issue, this study proposes a dual-topological learning framework that integrates topological constraints with geometric shape awareness. Specifically, a topological loss based on persistent homology will be combined with conventional segmentation objectives to ensure the preservation of connectivity and structural features. Additionally, a multi-channel neural network architecture will be utilized to enhance the representation of both global and local features. The expected outcome of this research is a robust and reliable segmentation method capable of maintaining both topology and geometry. This approach is anticipated to significantly improve segmentation performance in challenging scenarios involving complex structures. Furthermore, the results can be applied to various domains, including medical image analysis (e.g., vessel or tumor segmentation), GPR-based underground structure detection, and industrial inspection tasks such as defect detection. Through the collaboration at IMI, this research will be further refined and validated, contributing to the advancement of trustworthy and interpretable segmentation technologies. |
| Organizing Committee Members (Workshop) Participants (Short-term Joint Usage) |
Shizuo Kaji（Kyoto University・Professor） |