[Paper Review] A Voting-Stacking Ensemble of Inception Networks for Cervical Cytology Classification

Linyi Qian, Qian Huang+2 more|arXiv (Cornell University)|Aug 5, 2023

Cervical Cancer and HPV ResearchMedicine3 citations

TL;DR

This paper proposes a Voting-Stacking ensemble of three Inception networks for cervical cytology classification, leveraging misclassified samples from a voting ensemble to train a meta-learner, achieving 100% accuracy on SIPakMed, Herlev, and Mendeley datasets—surpassing state-of-the-art methods through improved feature fusion and robustness.

ABSTRACT

Cervical cancer is one of the most severe diseases threatening women's health. Early detection and diagnosis can significantly reduce cancer risk, in which cervical cytology classification is indispensable. Researchers have recently designed many networks for automated cervical cancer diagnosis, but the limited accuracy and bulky size of these individual models cannot meet practical application needs. To address this issue, we propose a Voting-Stacking ensemble strategy, which employs three Inception networks as base learners and integrates their outputs through a voting ensemble. The samples misclassified by the ensemble model generate a new training set on which a linear classification model is trained as the meta-learner and performs the final predictions. In addition, a multi-level Stacking ensemble framework is designed to improve performance further. The method is evaluated on the SIPakMed, Herlev, and Mendeley datasets, achieving accuracies of 100%, 100%, and 100%, respectively. The experimental results outperform the current state-of-the-art (SOTA) methods, demonstrating its potential for reducing screening workload and helping pathologists detect cervical cancer.

Motivation & Objective

• Address the limitations of single deep learning models in cervical cytology classification, including low accuracy and large model size.
• Reduce the screening workload for pathologists by enabling highly accurate, automated classification of cervical cell images.
• Overcome the poor generalization and feature redundancy in homogeneous models by introducing a novel ensemble strategy.
• Improve classification performance through a two-stage ensemble framework that selectively re-trains on misclassified samples.
• Demonstrate the method's robustness and scalability across diverse, publicly available cervical cytology datasets.

Proposed method

• Preprocess images via uniform resizing and apply combined online and offline data augmentation to enhance generalization.
• Utilize three pre-trained Inception family models (InceptionV3, InceptionResNetV2, Xception) as homogeneous base learners for feature extraction.
• Apply a voting strategy to aggregate predictions from base learners, identifying samples with conflicting outputs as candidates for meta-learning.
• Train a linear meta-learner exclusively on misclassified samples to refine predictions and reduce interference from correctly classified instances.
• Implement a multi-level Stacking ensemble framework by retraining the meta-learner on the outputs of the Voting-Stacking ensemble, further boosting performance.
• Employ transfer learning by initializing base models on ImageNet and fine-tuning on cervical cytology datasets to improve feature representation.

Experimental results

Research questions

• RQ1 Can a voting-stacking ensemble strategy improve classification accuracy in cervical cytology beyond single-model approaches?
• RQ2 Does selectively retraining a meta-learner on only misclassified samples enhance model performance compared to standard stacking?
• RQ3 How does the proposed Voting-Stacking framework compare to existing state-of-the-art methods on benchmark cervical cytology datasets?
• RQ4 To what extent does the multi-level Stacking framework further improve accuracy and robustness in cervical cell classification?
• RQ5 Can the proposed ensemble method be generalized to other medical image classification tasks beyond cervical cytology?

Key findings

• The proposed Voting-Stacking ensemble achieved 100% accuracy, precision, recall, and F1-score on the SIPakMed dataset, outperforming all prior state-of-the-art methods.
• On the Herlev dataset, the method achieved 100% accuracy across all evaluation metrics, demonstrating exceptional robustness and generalization.
• The Mendeley dataset evaluation confirmed 100% performance across all metrics, with the model significantly surpassing previous SOTA methods such as FuzzyDistanceEnsemble (99.68% accuracy) and ExemplarPyramid (99.47% accuracy).
• The 5-fold cross-validation results showed consistent performance across different training set sizes (10%, 20%, 30%), with the model maintaining 100% accuracy in all cases.
• The multi-level Stacking ensemble further improved accuracy by retraining the meta-learner on the outputs of the Voting-Stacking ensemble, confirming the effectiveness of hierarchical ensemble design.
• The method demonstrated superior performance even on challenging cell classes such as parabasal and abnormal cells, indicating strong discriminative capability across diverse cytological features.