MDFI: Multi-CNN Decision Feature Integration for Diagnosis of Cervical Precancerous Lesions

Yan-Min Luo, Tao Zhang, Ping Li, Peizhong Liu, Pengming Sun, BinHua Dong, Guanyu Ruan

Abstract

Cervical cancer remains a leading cause of mortality among women globally, primarily due to the subjective nature of colposcopy interpretation. This paper presents MDFI (Multi-CNN Decision Feature Integration), a deep learning framework designed to enhance the accuracy and sensitivity of cervical lesion diagnosis. By integrating K-means clustering for balanced data preprocessing and an ensemble of ResNet and DenseNet architectures fused via XGBoost, the MDFI system achieves a sensitivity of 0.704, significantly outperforming senior clinical experts and standalone CNN models.

Introduction

The standard "three-stage" technique for cervical cancer screening—cytology, colposcopy, and biopsy—relies heavily on the visual interpretation of colposcopy images. This step is inherently subjective, leading to high inter-observer variability and a significant risk of missed diagnoses, especially in low-resource regions. While deep learning has shown promise in medical imaging, single-model approaches often struggle with the subtle, diverse features of cervical lesions and the limited availability of high-quality labeled data.

Methods

The MDFI framework addresses these challenges through a three-stage diagnostic pipeline:

  1. K-means Preprocessing: To mitigate data imbalance and improve generalization, K-means clustering groups training images into $K$ classes based on feature similarity. This ensures that the 5-fold cross-validation sets are representative of the entire dataset's diversity.
  2. Transfer Learning: The system fine-tunes pre-trained ResNet50 and DenseNet121 architectures. These models are selected for their ability to extract both low-level textural details and high-level abstract global information.
  3. Multi-Level Integration: The core innovation is the integration strategy. Inner-model integration stabilizes the output of individual architectures across folds, while outer-model integration combines the strengths of different architectures. Finally, an XGBoost meta-classifier fuses these outputs to produce an optimized final prediction.

Datasets and Visual Complexity

The study utilized a clinical dataset of 1,709 cases collected from cooperating hospitals. Ground truth was established through the "gold standard" of biopsy results. The visual task is challenging due to the subtle differences between normal tissue and various grades of Cervical Intraepithelial Neoplasia (CIN).

Figure 5: Representative colposcopy images across diagnostic categories: Normal, CIN1, CIN2, and CIN3+.

Key Results

The MDFI framework demonstrated clear superiority over both standalone models and human experts:

  • Superior Sensitivity: The best MDFI strategy achieved a Sensitivity (SEN) of 0.704, significantly higher than the senior physician's score of 0.535 on the same test set. In clinical screening, high sensitivity is paramount to ensure no patient is left behind.
  • Improved AUC: The Area Under the Curve (AUC) reached 0.835, outperforming the baseline DenseNet (0.762) and ResNet (0.754) models.
  • Efficiency: The system can process 100 images in less than one minute, offering a massive speedup compared to manual clinical review.

Conclusion

MDFI provides a robust tool for computer-aided diagnosis, offering a more sensitive and objective alternative to traditional visual inspection. While the multi-stage training process is computationally intensive, the performance gains in sensitivity suggest it could serve as a vital "AI assistant" for clinicians. Future work will focus on validating the model across larger, multi-center datasets to further improve its robustness across different populations and imaging hardware.

Summary updated

Abstract

Colposcopy is an essential medical examination that mainly through visual inspection for cervical epithelial tissue to preventing cervical cancer. However, screening by artificial vision has the problems of missed diagnosis, low efficiency, and diversity. This paper proposes a deep learning-based method using multi-CNN decision feature integration for classification and diagnosis of cervical lesions. The proposed method first uses the k-means algorithm to aggregate training data into specific classes in data preprocessing and then trained in cross-validation to improve the generalization ability of the model. Subsequently, two kinds of CNN are separately fine-tuned based on the transfer learning, and then XGBoost algorithm is used to integrate the different CNN decision results for optimizing the final prediction. Two integration ways for decision features (named inner-model integration and outer-model integration) are proposed to embody multi-feature integration from the data level and the model structural level. Simultaneously, two specific integration strategies (inner-to-outer and outer-to-inner) are designed for as the final output based on the combination of the two ways. Experimental results show that the K-means data preprocessing method can improve the training of neural networks, and the proposed multi-decision feature fusion strategy can better obtain the results of computer-aided diagnosis that meet the needs of clinical diagnosis.
BibTeX
@article{luo2020mdfi,
  author = {Yan-Min Luo and Tao Zhang and Ping Li and Peizhong Liu and Pengming Sun and BinHua Dong and Guanyu Ruan},
  title = {{MDFI: Multi-CNN Decision Feature Integration for Diagnosis of Cervical Precancerous Lesions}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2972610},
}
Figure