BoNuS: Boundary Mining for Nuclei Segmentation with Partial Point Labels

TL;DR

BoNuS tackles nuclei instance segmentation with partial point labels by introducing boundary mining guided by a multiple-instance learning formulation. The method combines a detection module trained with Gaussian-encoded point annotations and curriculum learning, a coarse segmentation stage using Voronoi and cluster annotations, and a fine segmentation stage that learns pixel affinity to explicitly model boundaries. Key contributions include the boundary mining loss, the curriculum-driven pseudo-labeling, and the integration of coarse annotations to drive precise boundary-aware segmentation, achieving state-of-the-art or competitive results on MoNuSeg, CPM, and CoNIC. The approach significantly reduces annotation burden while delivering robust, instance-level nuclei segmentation suitable for downstream digital pathology analyses.

Abstract

Nuclei segmentation is a fundamental prerequisite in the digital pathology workflow. The development of automated methods for nuclei segmentation enables quantitative analysis of the wide existence and large variances in nuclei morphometry in histopathology images. However, manual annotation of tens of thousands of nuclei is tedious and time-consuming, which requires significant amount of human effort and domain-specific expertise. To alleviate this problem, in this paper, we propose a weakly-supervised nuclei segmentation method that only requires partial point labels of nuclei. Specifically, we propose a novel boundary mining framework for nuclei segmentation, named BoNuS, which simultaneously learns nuclei interior and boundary information from the point labels. To achieve this goal, we propose a novel boundary mining loss, which guides the model to learn the boundary information by exploring the pairwise pixel affinity in a multiple-instance learning manner. Then, we consider a more challenging problem, i.e., partial point label, where we propose a nuclei detection module with curriculum learning to detect the missing nuclei with prior morphological knowledge. The proposed method is validated on three public datasets, MoNuSeg, CPM, and CoNIC datasets. Experimental results demonstrate the superior performance of our method to the state-of-the-art weakly-supervised nuclei segmentation methods. Code: https://github.com/hust-linyi/bonus.

Figures (10)

• Figure 1: Illustration for boundary loss, where the boundary prediction can be supervised by pixel-affinity via multiple instance learning.
• Figure 2: This Figure illustrates the annotations at different granularities for nuclei image segmentation. The original image is shown in (a), followed by partial point annotations in (b) and full point annotations in (c). The distance transformation map is shown in (d), while Voronoi annotations and cluster annotations are synthesized from the full point annotations and displayed in (e) and (f), respectively. Pixel-wise binary annotations and instance annotations are shown in (g) and (h), respectively. It should be noted that the synthesized Voronoi and cluster annotations may be inaccurate and incomplete. The foreground, background, and ignore areas are highlighted in green, red, and black, respectively.
• Figure 3: An overview of the proposed method. The nuclei detection module with partial point label is shown in (a). This is followed by the coarse-stage nuclei segmentation module in (b), which is supervised by Voronoi and cluster labels derived from the detection prediction in the previous step. Finally, the fine-stage segmentation module in (c) produces binary segmentation results that are supervised by the coarse segmentation results, while the boundary prediction is supervised by the affinity map in a multiple instance learning manner. The dashed lines denote the supervision information.
• Figure 4: Architectures of (a) coarse-stage segmentation model; (b) fine-stage segmentation model.
• Figure 5: The boundary loss used in our method. The left panel shows how the nuclei instance prediction of the coarse model is used to supervise the boundary prediction of the fine-stage segmentation model. The right panel illustrates the calculation of the boundary loss using pairwise pixel affinity.