Robust Whole Slide Image Analysis for Cervical Cancer Screening Using Deep Learning

Shenghua Cheng, Sibo Liu, Jingya Yu, Gong Rao, Yuwei Xiao, Wei Han, Wenjie Zhu, Xiaohua Lv, Ning Li, Jing Cai, Zehua Wang, Feng Xi, Fei Yang, Xiebo Geng, Jiabo Ma, Li Xu, Ziquan Wei, Xueying Zhang, Shaoqun Zeng, Li Chen, Junbo Hu, Xiuli Liu

Introduction

Cervical cancer screening is a critical public health intervention, yet it remains heavily dependent on the manual review of Pap smears. This process is labor-intensive, subjective, and susceptible to human error due to fatigue. While digital pathology and artificial intelligence (AI) offer promising solutions, existing algorithms often struggle with the "Generalization Gap." AI models trained on data from a single institution frequently underperform when applied to slides from different hospitals, which may use varying staining techniques, slide preparation methods (e.g., liquid-based vs. sedimentation), and diverse digital scanners. Current commercial "closed-loop" systems often require proprietary hardware, creating a barrier for resource-limited regions. This study presents a robust, progressive deep learning system designed to overcome these challenges by mimicking the workflow of human pathologists.

Methodology: The Progressive Recognition System

The researchers developed a three-stage Progressive Recognition System that balances computational efficiency with high diagnostic accuracy:

  1. Low-Resolution (LR) Screening: A modified ResNet50 architecture analyzes the entire Whole Slide Image (WSI) at a coarse scale (0.486 µm/pixel). A semantic segmentation branch identifies suspicious regions, generating a heatmap of potential lesions.
  2. High-Resolution (HR) Identification: The system extracts high-resolution crops (0.243 µm/pixel) from the most suspicious areas identified in the LR stage. This model identifies specific lesion cells and ranks them by their probability of malignancy.
  3. RNN-Based Classification: To provide a slide-level diagnosis, a Recurrent Neural Network (RNN) aggregates features from the top 10 most suspicious cells. This approach addresses the "needle in a haystack" problem by focusing on the most abnormal findings rather than averaging data across the entire slide.

To ensure robustness, the team employed Diverse Data Learning, training the models on multiple styles simultaneously. They also utilized Hard Sample Mining to improve performance on difficult cases and extensive Data Augmentation to simulate variations in color and contrast.

Data and Validation

The system was trained and validated on a massive, multi-center dataset:

  • Volume: 3,545 patient-wise WSIs containing 79,218 expert-annotated lesion cells.
  • Diversity: Data was sourced from five different hospitals using five distinct digital scanner models.
  • Testing: The dataset was divided into 12 groups. Groups A-D were used for training, while Groups E-L served as independent test sets (comprising 1,170 patient-wise WSIs) to evaluate the system's ability to generalize to unseen clinical environments.

Results

The system demonstrated performance comparable to expert cytopathologists:

  • Diagnostic Accuracy: On the independent test sets, the model achieved 95.1% Sensitivity and 93.5% Specificity, placing it on par with the average performance of three independent expert cytopathologists.
  • Cell Recommendation: The True Positive Rate (TPR) for recommending the top 10 lesion cells was 88.5% on 447 positive slides, significantly outperforming the industry-standard Hologic ThinPrep Imaging System (TIS).
  • Operational Efficiency: A giga-pixel WSI can be processed in approximately 1.5 minutes using a single consumer-grade GPU (Nvidia 1080Ti).
  • Workload Reduction: The system successfully excluded approximately 49% of negative slides with 100% sensitivity, suggesting it could halve the screening workload for pathologists without missing a single positive case.

Discussion and Future Work

By mimicking the "low-power search, high-power confirmation" workflow used by human experts, this progressive system achieves high accuracy and speed. t-SNE visualizations confirmed that the training strategies helped the network learn "universal" features of cervical lesions that are invariant to staining and scanning differences.

While the current system is robust, it currently relies on professional-grade digital scanners. Future research will focus on adapting these algorithms for use with low-cost, portable hardware and augmented reality (AR) microscopes that can highlight suspicious cells in real-time. Additionally, the researchers plan to expand the model's capabilities to include subclass classification (e.g., distinguishing between LSIL and HSIL), addressing the challenges of high inter-observer variability in clinical practice.

Summary updated

Abstract

Abstract Computer-assisted diagnosis is key for popularizing cervical cancer screening. However, current recognition algorithms are insufficient in accuracy and generalization for cervical lesion cells, especially when facing diversity data in clinical applications. Inspired by manual reading slide under microscopes, we develop a progressive lesion cell recognition method combing low and high resolutions WSIs to recommend lesion cells and a recurrent neural network-based WSI classification model to evaluate the lesion degree of WSIs. After validating our system on 3,545 patient-wise WSIs with 79,218 annotations from multiple hospitals and several imaging instruments, on multi-center independent test sets of 1,170 patient-wise WSIs, we achieve 93.5% Specificity and 95.1% Sensitivity for classifying slides, closely equivalent to the average level of three independent cytopathologists, and obtain 88.5% TPR (true positive rate) for recommending top 10 lesion cells on 447 positive slides. After deploying, our system recognizes one giga-pixel WSI in about 1.5 minutes using one Nvidia 1080Ti GPU.
BibTeX
@inproceedings{cheng2021robust,
  author = {Shenghua Cheng and Sibo Liu and Jingya Yu and Gong Rao and Yuwei Xiao and Wei Han and Wenjie Zhu and Xiaohua Lv and Ning Li and Jing Cai and Zehua Wang and Feng Xi and Fei Yang and Xiebo Geng and Jiabo Ma and Li Xu and Ziquan Wei and Xueying Zhang and Shaoqun Zeng and Li Chen and Junbo Hu and Xiuli Liu},
  title = {{Robust Whole Slide Image Analysis for Cervical Cancer Screening Using Deep Learning}},
  booktitle = {Nature Communications},
  year = {2021},
  doi = {10.1038/s41467-021-25296-x},
}
Figure