SEINE: Structure Encoding and Interaction Network for Nuclei Instance Segmentation.

Area of Science:

• Digital pathology
• Computational biology
• Medical image analysis

Background:

• Nuclei instance segmentation is vital for cancer diagnosis and biological studies.
• Challenges include under-segmentation due to poor staining and fragmented predictions from nuclear textures.

Purpose of the Study:

• To propose a new network, SEINE, for accurate and integral nuclei instance segmentation.
• To address under-segmentation and fragmentation issues in histopathological images.

Main Methods:

• Developed a Structure Encoding and Interaction NEtwork (SEINE).
• Introduced a contour-based structure encoding mechanism for accurate structural representation.
• Implemented a structure-guided attention module and a position enhancement strategy with centroid distance constraint.

Main Results:

• SEINE effectively addresses under-segmentation by using clear nuclei to guide unclear ones.
• The position enhancement strategy reduces contour prediction errors, mitigating fragmentation.
• Achieved state-of-the-art performance on four benchmark datasets.

Conclusions:

• SEINE demonstrates superior performance in nuclei instance segmentation.
• The proposed methods effectively overcome common challenges in histopathological image analysis.
• The approach holds significant potential for improving cancer diagnosis and biological analysis.