Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Improved Cervical Cancer Screening: Vision Transformer-Based Classification and Interpretability

Khoa Tuan Nguyen, Ho-min Park, Gaeun Oh, Joris Vankerschaver, Wesley De Neve

TL;DR

The paper tackles automated cervical cancer screening by classifying Pap smear images into Healthy, Unhealthy, and Rubbish categories using a fine-tuned EVA-02 Vision Transformer. It introduces a four-step pipeline—feature extraction from transformer tokens, feature selection via classical models, training a deeper ANN with optional loss weighting, and Kernel SHAP-based interpretability—and demonstrates a peak F1-score of $0.85227$, outperforming the EVA-02 baseline of $0.84878$. The work also highlights interpretability by identifying feature #10 as most influential and linking it to cell morphology and staining patterns. While the results show a small but consistent improvement and enhanced explainability, the evaluation is limited to the public leaderboard, underscoring the need for broader validation in practical screening settings.

Abstract

We propose a novel approach to cervical cell image classification for cervical cancer screening using the EVA-02 transformer model. We developed a four-step pipeline: fine-tuning EVA-02, feature extraction, selecting important features through multiple machine learning models, and training a new artificial neural network with optional loss weighting for improved generalization. With this design, our best model achieved an F1-score of 0.85227, outperforming the baseline EVA-02 model (0.84878). We also utilized Kernel SHAP analysis and identified key features correlating with cell morphology and staining characteristics, providing interpretable insights into the decision-making process of the fine-tuned model. Our code is available at https://github.com/Khoa-NT/isbi2025_ps3c.

Towards Improved Cervical Cancer Screening: Vision Transformer-Based Classification and Interpretability

TL;DR

The paper tackles automated cervical cancer screening by classifying Pap smear images into Healthy, Unhealthy, and Rubbish categories using a fine-tuned EVA-02 Vision Transformer. It introduces a four-step pipeline—feature extraction from transformer tokens, feature selection via classical models, training a deeper ANN with optional loss weighting, and Kernel SHAP-based interpretability—and demonstrates a peak F1-score of , outperforming the EVA-02 baseline of . The work also highlights interpretability by identifying feature #10 as most influential and linking it to cell morphology and staining patterns. While the results show a small but consistent improvement and enhanced explainability, the evaluation is limited to the public leaderboard, underscoring the need for broader validation in practical screening settings.

Abstract

We propose a novel approach to cervical cell image classification for cervical cancer screening using the EVA-02 transformer model. We developed a four-step pipeline: fine-tuning EVA-02, feature extraction, selecting important features through multiple machine learning models, and training a new artificial neural network with optional loss weighting for improved generalization. With this design, our best model achieved an F1-score of 0.85227, outperforming the baseline EVA-02 model (0.84878). We also utilized Kernel SHAP analysis and identified key features correlating with cell morphology and staining characteristics, providing interpretable insights into the decision-making process of the fine-tuned model. Our code is available at https://github.com/Khoa-NT/isbi2025_ps3c.
Paper Structure (14 sections, 2 equations, 3 figures, 2 tables)

This paper contains 14 sections, 2 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Fine-tuning EVA-02
  4. (A) Feature extraction
  5. (B) Feature selection
  6. (C) Training a new ANN model
  7. (D) SHAP-based feature analysis
  8. Results and discussion
  9. (A) Impact of feature extraction
  10. (B) Feature selection effectiveness
  11. (C) Effect of loss weighting
  12. Comparison with baseline EVA-02
  13. (D) SHAP-based feature analysis
  14. Conclusion

Figures (3)

  • Figure 1: A simple illustrative workflow of the EVA-02 model. The input cell image is divided into patches, creating Image tokens. A sequence of transformer-based blocks processes a trainable Class token along with these Image tokens. The output Image tokens are then average pooled before being fed into the classifier to predict Healthy, Unhealthy, and Rubbish classes. The Class token is extracted at (1), while the Image tokens are extracted at (2).
  • Figure 2: Overview of the experiment.
  • Figure 3: Comparison of cell images based on feature #10 values, demonstrating distinct morphological and staining characteristics.