SEINE: Structure Encoding and Interaction Network for Nuclei Instance Segmentation

TL;DR

SEINE tackles nuclei instance segmentation with chromophobe nuclei by introducing a contour-based structure encoding that links semantic and structural information. A Structure Guided Attention module enables cross-nuclei structure interaction, while Semantic Feature Fusion and Position Enhancement improve feature alignment and boundary integrity. The framework is reinforced by post-processing to fuse structure cues into semantic predictions. Across four datasets, SEINE achieves state-of-the-art results, demonstrating strong robustness to structural variations and rotations; code is publicly available.

Abstract

Nuclei instance segmentation in histopathological images is of great importance for biological analysis and cancer diagnosis but remains challenging for two reasons. (1) Similar visual presentation of intranuclear and extranuclear regions of chromophobe nuclei often causes under-segmentation, and (2) current methods lack the exploration of nuclei structure, resulting in fragmented instance predictions. To address these problems, this paper proposes a structure encoding and interaction network, termed SEINE, which develops the structure modeling scheme of nuclei and exploits the structure similarity between nuclei to improve the integrality of each segmented instance. Concretely, SEINE introduces a contour-based structure encoding (SE) that considers the correlation between nuclei structure and semantics, realizing a reasonable representation of the nuclei structure. Based on the encoding, we propose a structure-guided attention (SGA) module that takes the clear nuclei as prototypes to enhance the structure learning for the fuzzy nuclei. To strengthen the structural learning ability, a semantic feature fusion (SFF) is presented to boost the semantic consistency of semantic and structure branches. Furthermore, a position enhancement (PE) method is applied to suppress incorrect nuclei boundary predictions. Extensive experiments demonstrate the superiority of our approaches, and SEINE achieves state-of-the-art (SOTA) performance on four datasets. The code is available at https://github.com/zhangye-zoe/SEINE.

Figures (12)

• Figure 1: Comparison of structure modeling methods. (a) is the input image, (b) is the horizontal distance encoding, (c) is the centripetal direction encoding, and (d) is our proposed structure encoding (SE), whose contour is highlighted in the green line. The second row represents the structure modeling when using rotation data augmentation.
• Figure 2: The comparison of the general methods and our SEINE. The third row exhibits a chromophobe and a clearly stained nucleus. In the chromophobe, the intranuclear region is ambiguous, but the membrane is distinct. In the clearly stained nucleus, both the intranuclear region and membrane are clear. The double-headed arrow in (b) represents the structure interaction.
• Figure 3: The flowchart of the proposed SEINE. The top and the bottom of the flowchart represent training phase and testing phase, repectively. In the training phase, the network contains a shared feature extractor, a semantic branch, and a structure branch. For each block of the two branches, the semantic feature fusion (SFF) module and the structure-guided attention (SGA) module are constructed to perform structure interaction. The position enhancement (PE) is used to enhance the localization ability of nuclei centroid. In the testing phase, the prediction results from the semantic branch and structure branch are combined for refined instance results.
• Figure 4: The optimization procedure of the semantic feature and structure feature between two sequential blocks. The feature outputs of SFF are fed into SGA for structure interactions.
• Figure 5: The illustration of structure-guided attention module. In the SGA map, the "In.", "Co.", and "Ex." representents the intranuclear, contour, and extranuclear regions, respectively.