Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Holistic and Historical Instance Comparison for Cervical Cell Detection

Hao Jiang, Runsheng Liu, Yanning Zhou, Huangjing Lin, Hao Chen

TL;DR

This work tackles cervical cell detection in Pap smear cytology, where subtle class ambiguity and significant class imbalance hinder accurate abnormal-cell identification. It introduces a holistic and historical instance comparison framework comprising RoI-level contrastive learning (RIC) and class-level contrastive learning (CIC) with a confident-sample memory bank, enabling stronger discriminability across both foreground regions and class categories. Built on a Faster R-CNN backbone, the approach uses box augmentation and memory-bank sampling to boost minority-class representation and applies joint losses to align RoI and class embeddings. Experiments on two large-scale datasets (CC-L and CC-S) demonstrate marked improvements in AP50, AP75, AP, and AR, with substantial gains for ambiguous and minority classes, and provide evidence of practical impact for cervical cancer screening workflows.

Abstract

Cytology screening from Papanicolaou (Pap) smears is a common and effective tool for the preventive clinical management of cervical cancer, where abnormal cell detection from whole slide images serves as the foundation for reporting cervical cytology. However, cervical cell detection remains challenging due to 1) hazily-defined cell types (e.g., ASC-US) with subtle morphological discrepancies caused by the dynamic cancerization process, i.e., cell class ambiguity, and 2) imbalanced class distributions of clinical data may cause missed detection, especially for minor categories, i.e., cell class imbalance. To this end, we propose a holistic and historical instance comparison approach for cervical cell detection. Specifically, we first develop a holistic instance comparison scheme enforcing both RoI-level and class-level cell discrimination. This coarse-to-fine cell comparison encourages the model to learn foreground-distinguishable and class-wise representations. To emphatically improve the distinguishability of minor classes, we then introduce a historical instance comparison scheme with a confident sample selection-based memory bank, which involves comparing current embeddings with historical embeddings for better cell instance discrimination. Extensive experiments and analysis on two large-scale cytology datasets including 42,592 and 114,513 cervical cells demonstrate the effectiveness of our method. The code is available at https://github.com/hjiangaz/HERO.

Holistic and Historical Instance Comparison for Cervical Cell Detection

TL;DR

This work tackles cervical cell detection in Pap smear cytology, where subtle class ambiguity and significant class imbalance hinder accurate abnormal-cell identification. It introduces a holistic and historical instance comparison framework comprising RoI-level contrastive learning (RIC) and class-level contrastive learning (CIC) with a confident-sample memory bank, enabling stronger discriminability across both foreground regions and class categories. Built on a Faster R-CNN backbone, the approach uses box augmentation and memory-bank sampling to boost minority-class representation and applies joint losses to align RoI and class embeddings. Experiments on two large-scale datasets (CC-L and CC-S) demonstrate marked improvements in AP50, AP75, AP, and AR, with substantial gains for ambiguous and minority classes, and provide evidence of practical impact for cervical cancer screening workflows.

Abstract

Cytology screening from Papanicolaou (Pap) smears is a common and effective tool for the preventive clinical management of cervical cancer, where abnormal cell detection from whole slide images serves as the foundation for reporting cervical cytology. However, cervical cell detection remains challenging due to 1) hazily-defined cell types (e.g., ASC-US) with subtle morphological discrepancies caused by the dynamic cancerization process, i.e., cell class ambiguity, and 2) imbalanced class distributions of clinical data may cause missed detection, especially for minor categories, i.e., cell class imbalance. To this end, we propose a holistic and historical instance comparison approach for cervical cell detection. Specifically, we first develop a holistic instance comparison scheme enforcing both RoI-level and class-level cell discrimination. This coarse-to-fine cell comparison encourages the model to learn foreground-distinguishable and class-wise representations. To emphatically improve the distinguishability of minor classes, we then introduce a historical instance comparison scheme with a confident sample selection-based memory bank, which involves comparing current embeddings with historical embeddings for better cell instance discrimination. Extensive experiments and analysis on two large-scale cytology datasets including 42,592 and 114,513 cervical cells demonstrate the effectiveness of our method. The code is available at https://github.com/hjiangaz/HERO.
Paper Structure (12 sections, 5 equations, 3 figures, 6 tables)

This paper contains 12 sections, 5 equations, 3 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Framework Overview
  4. Holistic Instance Comparison
  5. Historical Instance Comparison
  6. Overall Loss Function and Optimization
  7. Experiments and Results
  8. Dataset and Experiments Settings
  9. Comparison with SOTA methods
  10. Ablation Studies and Analysis
  11. Conclusion
  12. Acknowledgments

Figures (3)

  • Figure 1: Illustration of cervical cell detection. (a) Detected abnormal cells denoted by blue boxes; (b) Cell class imbalance: class imbalanced cell instance distribution; (c) Cell class ambiguity: cell class ambiguity with various appearances and morphologies.
  • Figure 2: Overview of the proposed Holistic and Historical Instance Comparison framework (a) for instance comparison (b). It consists of contrasting RoI features using RoI-level instance comparison module (RIC) with box augmentation (c), and contrasting class features by class-level comparison module (CIC) with a confident sample selection-based memory bank (d).
  • Figure 3: Qualitative results of our methods and other SOTA methods. (a) FRCNN, (b) Cascade R-CNN, (c) Grid R-CNN, (d) Sparse R-CNN, (e) RepPoints, (f) CornerNet, (g) YOLOv3, (h) RetinaNet, (i) FCOS, (j) DETR, (k) LOCE, our method with R50 (l) and R101 (m), and (n) GT.