Comprehensive Attention Networks for Cell Segmentation Using Numerical Integration
| Authors | |
|---|---|
| Year of publication | 2026 |
| Type | Paper in proceedings |
| Conference | 2026 IEEE International Symposium on Biomedical Imaging (ISBI) |
| MU Faculty or unit | |
| Citation | |
| Keywords | Image processing; Biomedical image analysis |
| Attached files | |
| Description | The emergence of neural networks has greatly improved segmentation accuracy in the domain of biomedical imaging, but the majority of existing solutions all employ the same general methodology. Attempting to generate entire masks at once naturally makes it difficult for a network to perform consistently well in every boundary region, especially when the objects to be segmented are complex. We describe a novel approach which combines attention-based neural networks with numerical integration methods to segment detected objects in a piecewise manner by tracing boundary pixels. This algorithm allows datasets with only a small number of annotated instances to be used for training and for behaviour analysis at the inference stage. We demonstrate our results on two datasets of the Cell Tracking Challenge, including simulated and real recordings of cells. The source code is available at https://gitlab.fi.muni.cz/xeftimiu/limit-cycle-seg. |
| Related projects: |