Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Annotation-Efficient Polyp Segmentation via Active Learning

Duojun Huang, Xinyu Xiong, De-Jun Fan, Feng Gao, Xiao-Jian Wu, Guanbin Li

TL;DR

The paper tackles the annotation burden in polyp segmentation by introducing an annotation-efficient deep active learning framework. It jointly leverages an uncertain-weighted clustering of image-level features to select $B$ informative unlabeled samples per round based on $I(x^u) = \cos(F_1(x^u), F_0(x^u))$, and a novel unsupervised feature discrepancy loss $\mathcal{L}_{fdl}$ to sharpen class separation on unlabeled data. Training uses a supervised loss $\mathcal{L}_{seg}$ plus $\lambda_c \mathcal{L}_{fdl}$ with $\lambda_c=0.1$, and employs $R=5$ rounds with $B=30$ samples per round. Experiments on CVC-ClinicDB and an in-house dataset show state-of-the-art performance under limited annotation budgets, with gains of about 1–2 percentage points in $mIoU$ and Dice over strong baselines, demonstrating reduced labeling burden and improved generalization for colonoscopy polyp segmentation.

Abstract

Deep learning-based techniques have proven effective in polyp segmentation tasks when provided with sufficient pixel-wise labeled data. However, the high cost of manual annotation has created a bottleneck for model generalization. To minimize annotation costs, we propose a deep active learning framework for annotation-efficient polyp segmentation. In practice, we measure the uncertainty of each sample by examining the similarity between features masked by the prediction map of the polyp and the background area. Since the segmentation model tends to perform weak in samples with indistinguishable features of foreground and background areas, uncertainty sampling facilitates the fitting of under-learning data. Furthermore, clustering image-level features weighted by uncertainty identify samples that are both uncertain and representative. To enhance the selectivity of the active selection strategy, we propose a novel unsupervised feature discrepancy learning mechanism. The selection strategy and feature optimization work in tandem to achieve optimal performance with a limited annotation budget. Extensive experimental results have demonstrated that our proposed method achieved state-of-the-art performance compared to other competitors on both a public dataset and a large-scale in-house dataset.

Annotation-Efficient Polyp Segmentation via Active Learning

TL;DR

The paper tackles the annotation burden in polyp segmentation by introducing an annotation-efficient deep active learning framework. It jointly leverages an uncertain-weighted clustering of image-level features to select informative unlabeled samples per round based on , and a novel unsupervised feature discrepancy loss to sharpen class separation on unlabeled data. Training uses a supervised loss plus with , and employs rounds with samples per round. Experiments on CVC-ClinicDB and an in-house dataset show state-of-the-art performance under limited annotation budgets, with gains of about 1–2 percentage points in and Dice over strong baselines, demonstrating reduced labeling burden and improved generalization for colonoscopy polyp segmentation.

Abstract

Deep learning-based techniques have proven effective in polyp segmentation tasks when provided with sufficient pixel-wise labeled data. However, the high cost of manual annotation has created a bottleneck for model generalization. To minimize annotation costs, we propose a deep active learning framework for annotation-efficient polyp segmentation. In practice, we measure the uncertainty of each sample by examining the similarity between features masked by the prediction map of the polyp and the background area. Since the segmentation model tends to perform weak in samples with indistinguishable features of foreground and background areas, uncertainty sampling facilitates the fitting of under-learning data. Furthermore, clustering image-level features weighted by uncertainty identify samples that are both uncertain and representative. To enhance the selectivity of the active selection strategy, we propose a novel unsupervised feature discrepancy learning mechanism. The selection strategy and feature optimization work in tandem to achieve optimal performance with a limited annotation budget. Extensive experimental results have demonstrated that our proposed method achieved state-of-the-art performance compared to other competitors on both a public dataset and a large-scale in-house dataset.
Paper Structure (11 sections, 7 equations, 3 figures, 2 tables)

This paper contains 11 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Sampling Query Strategy
  4. Feature Discrepancy Learning
  5. Loss Function
  6. Experiments
  7. Experimental Setup
  8. Experimental Results
  9. Conclusion
  10. Compliance with Ethical Standards
  11. Acknowledgments

Figures (3)

  • Figure 1: The work flow of our proposed active learning framework. Given an unlabeled sample $x^u$ from the unlabeled set $D^U$, our goal is to accurately measure the value of $x^u$, and high value one will be selected for expert annotation.
  • Figure 2: Comparison between the a) traditional pure K-means and b) our proposed uncertain-weighted clustering algorithm.
  • Figure 3: Performance comparison between different method on CVC-ClinicDB dataset. For simplicity, the full performance plots for the anonymous dataset are provided in the supplemental material.