Related Experiment Video
Updated: May 14, 2026
10:39
A Label-Free Segmentation Approach for Intravital Imaging of Mammary Tumor Microenvironment
Published on: May 24, 2022
Self-Refining Segment Anything Model for Nuclei Segmentation as Contrastive Learning Approach to Label-Efficient
1Department of Robotics and Mechatronics Engineering, Daegu Gyeongbuk Institute of Science and Technology (DGIST), Daegu 42988, Republic of Korea.
Diagnostics (Basel, Switzerland)
|May 13, 2026
Summary
This study introduces a self-evolving framework for nuclei instance segmentation using sparse point annotations, enhancing digital pathology accuracy. The method refines learning targets adaptively, achieving state-of-the-art results for automated diagnostics.
Area of Science:
- Digital Pathology
- Computational Biology
- Medical Image Analysis
Background:
- Precise nuclei instance segmentation is crucial for digital pathology.
- Scarcity of pixel-level annotations limits deep learning model performance.
Purpose of the Study:
- Develop a robust nuclei segmentation framework using sparse point annotations.
- Extend the Segment Anything Model (SAM) for histopathology applications.
- Improve automated diagnostic workflows in clinical settings.
Main Methods:
- Propose a self-evolving framework utilizing sparse point annotations.
- Implement a self-evolving labeling strategy with Exponential Moving Average (EMA) for adaptive target refinement.
- Integrate instance-aware contrastive learning and consensus-based filtering.
Main Results:
- Achieve state-of-the-art performance on CPM17, MoNuSeg, and CoNSeP datasets.
- Demonstrate high accuracy across various backbones (ViT-B, ViT-H).
- Validate the framework's effectiveness for nuclei instance segmentation.
Conclusions:
- The self-refining approach transitions foundation models to specialized histopathology tools.
- Offers an efficient and accurate solution for automated diagnostic workflows.
- Enables reliable digital pathology through precise nuclei segmentation.
