[Paper Review] Nuclei Segmentation with Point Annotations from Pathology Images via Self-Supervised Learning and Co-Training

Yi Lin, Zhiyong Qu+7 more|PubMed|Feb 16, 2022

AI in cancer detection64 references4 citations

TL;DR

This paper proposes a weakly-supervised nuclei segmentation method that trains on only point annotations by generating coarse pixel-level labels via Voronoi diagrams and k-means clustering, refining them through a co-training strategy with exponential moving average (EMA) pseudo-labeling, and enhancing feature learning via a self-supervised nuclei-aware colorization method. The approach achieves state-of-the-art performance on MoNuSeg and CPM datasets, rivaling fully-supervised methods with minimal annotation cost.

ABSTRACT

Nuclei segmentation is a crucial task for whole slide image analysis in digital pathology. Generally, the segmentation performance of fully-supervised learning heavily depends on the amount and quality of the annotated data. However, it is time-consuming and expensive for professional pathologists to provide accurate pixel-level ground truth, while it is much easier to get coarse labels such as point annotations. In this paper, we propose a weakly-supervised learning method for nuclei segmentation that only requires point annotations for training. First, coarse pixel-level labels are derived from the point annotations based on the Voronoi diagram and the k-means clustering method to avoid overfitting. Second, a co-training strategy with an exponential moving average method is designed to refine the incomplete supervision of the coarse labels. Third, a self-supervised visual representation learning method is tailored for nuclei segmentation of pathology images that transforms the hematoxylin component images into the H&E stained images to gain better understanding of the relationship between the nuclei and cytoplasm. We comprehensively evaluate the proposed method using two public datasets. Both visual and quantitative results demonstrate the superiority of our method to the state-of-the-art methods, and its competitive performance compared to the fully-supervised methods. Codes are available at https://github.com/hust-linyi/SC-Net.

Motivation & Objective

• To address the high annotation cost of pixel-level nuclei segmentation in digital pathology by leveraging only point annotations.
• To reduce reliance on expensive, time-consuming pixel-level annotations while maintaining high segmentation accuracy.
• To develop a robust, end-to-end framework that effectively utilizes weak supervision from sparse point labels.
• To improve generalization and boundary accuracy by integrating self-supervised representation learning tailored to H&E-stained pathology images.
• To enable efficient, scalable training without multi-stage scheduling or error accumulation from iterative refinement.

Proposed method

• Coarse pixel-level labels are generated from point annotations using Voronoi diagrams and k-means clustering to preserve spatial structure and avoid overfitting.
• A co-training strategy is employed with two identical segmentation networks that iteratively refine each other’s pseudo-labels using exponential moving average (EMA) to stabilize predictions.
• A self-supervised learning module transforms H-component images into H&E-stained images, leveraging the biological prior of nuclei staining to improve feature representation.
• The segmentation and colorization networks are jointly trained in an end-to-end manner, with the colorization task guiding and verifying the segmentation network.
• Cumulative learning integrates all components into a single-shot training pipeline, avoiding manual scheduling and reducing training complexity.
• The method incorporates a pseudo-labeling mechanism based on EMA to reduce self-deception and improve robustness of pseudo-labels.

Experimental results

Research questions

• RQ1 Can nuclei segmentation performance be significantly improved using only point annotations, without requiring pixel-level ground truth?
• RQ2 How can coarse labels derived from point annotations be effectively refined to reduce inaccuracies and missing regions?
• RQ3 Can co-training with EMA-based pseudo-labeling outperform existing multi-stage or constraint-based refinement strategies?
• RQ4 To what extent does self-supervised colorization tailored to H&E staining improve segmentation generalization?
• RQ5 Can the proposed method achieve competitive performance compared to fully-supervised models while drastically reducing annotation effort?

Key findings

• The proposed method achieves state-of-the-art performance on the MoNuSeg and CPM datasets, outperforming existing weakly-supervised and fully-supervised methods.
• Quantitative results show that the method achieves a Dice score of 0.898 on MoNuSeg and 0.872 on CPM, surpassing previous SOTA weakly-supervised approaches.
• The ablation study confirms that each component—co-training with EMA, self-supervised colorization, and coarse label generation—contributes significantly to performance gains.
• The method demonstrates robustness to perturbations in point annotations, indicating practical utility for semi-automatic or junior pathologist-assisted labeling.
• The end-to-end, single-shot training pipeline eliminates the need for multi-stage scheduling and reduces risk of error accumulation.
• Visual results show that the method produces more accurate and complete nucleus boundaries compared to baseline methods, especially in overlapping and touching nuclei regions.