AI-Assisted Cervical Cytology Screening: Improving Accuracy and Efficiency with Digital Pathology

This article is from Chinese Clinical Oncology

Author: Liu Sichun, Hu huaiyuan, Zhu nana, and Tangheng

Why AI Matters in Cervical Cytology Screening

Cervical cancer remains one of the leading causes of cancer-related deaths among women worldwide. It is the fourth most common malignant tumor in women globally. In China alone, 109,741 new cervical cancer cases and 59,060 related deaths were reported in 2021. As a result, cervical cancer ranked as the eighth leading cause of cancer-related deaths among Chinese women.

Early detection is the most effective way to reduce cervical cancer mortality. Therefore, expanding screening coverage, adopting innovative screening models, and improving screening quality are critical. With effective screening programs, the early detection rate can exceed 90%.

However, traditional cervical cytology screening faces several challenges. China continues to experience a shortage of pathologists, resulting in heavy workloads and increasing diagnostic pressure. Even if one pathologist reviews 100 slides per day, the current workforce still cannot meet clinical demand. In addition, medical resources are unevenly distributed across different regions, and cytology quality control standards vary significantly.

Manual slide interpretation also depends heavily on the experience and condition of individual pathologists. Fatigue, differences in expertise, and subjective interpretation can reduce diagnostic sensitivity to approximately 65%. Consequently, both missed diagnoses and false diagnoses remain important concerns in routine cervical cytology screening.

How AI Improves Cervical Cytology Screening

To address these challenges, artificial intelligence (AI) has become an important technology in digital pathology. AI-assisted cervical cytology combines advanced image analysis with pathologists’ expertise to improve screening efficiency and diagnostic performance.

Instead of replacing pathologists, AI automatically pre-screens whole slide images and highlights suspicious abnormal cells for further review. As a result, pathologists can focus on clinically significant cases while reducing the time required for routine slide screening. This workflow improves information processing efficiency and supports more consistent cytological evaluation.

In this study, we compared an AI-assisted cervical cytology system with conventional manual microscopy. Specifically, we evaluated diagnostic sensitivity, specificity, and screening efficiency to assess the clinical value of AI-assisted slide interpretation in cervical cytology.

Clinical Study Design

To evaluate the clinical performance of AI-assisted cervical cytology, we compared conventional manual slide interpretation with an AI-assisted screening workflow. The study focused on diagnostic accuracy, sensitivity, specificity, and screening efficiency.

Study Population

The control group included 5,400 ThinPrep Cytologic Test (TCT) slides. All slides had been manually reviewed, diagnosed, and confirmed to meet quality requirements. Huikang Medical Laboratory collected these slides between January and August 2022.

The AI group included 4,893 slides that successfully passed the AI quality assessment and were considered valid for analysis.

Diagnostic Criteria

The study followed the 2014 Bethesda System (TBS) diagnostic criteria. Cytology results of ASC-US and above were classified as positive, while all other categories were classified as negative.

The evaluated diagnostic categories included:

• Negative for intraepithelial lesion or malignancy (NILM)
• Atypical squamous cells of undetermined significance (ASC-US)
• Atypical squamous cells—cannot exclude HSIL (ASC-H)
• Low-grade squamous intraepithelial lesion (LSIL)
• High-grade squamous intraepithelial lesion (HSIL)
• Squamous cell carcinoma (SCC)
• Atypical glandular cells (AGC)
• Adenocarcinoma in situ (AIS)
• Adenocarcinoma

Because glandular epithelial lesions were rare during the study period, only five AGC cases were reported. Therefore, glandular lesions were excluded from the final statistical analysis.

AI-Assisted Screening Workflow

The laboratory prepared all cytology slides using an automated sedimentation-based slide preparation system.

Next, four experienced pathologists—including two attending physicians and two associate chief physicians—performed conventional manual slide interpretation.

For the AI-assisted workflow, the KFBIO KF-PRO-400 Digital Pathology Scanner digitized every slide into high-resolution whole slide images (WSIs). The KFBIO AI Cervical Cytology Diagnosis System then automatically analyzed each image and identified suspicious abnormal cells for pathologist review.

This human-AI collaborative workflow combined automated image analysis with expert verification. As a result, it improved screening efficiency while maintaining diagnostic quality.

Results

AI Successfully Processed Most Cervical Cytology Slides

The study included 5,400 cervical cytology slides for AI analysis. Among them, the AI system successfully identified 4,893 slides as valid for evaluation.

The valid slides included:

• 4,094 NILM cases
• 20 fungal infection cases
• 5 trichomonas infection cases
• 280 ASC-US cases
• 267 LSIL cases
• 67 ASC-H cases
• 158 HSIL cases
• 2 SCC cases

The remaining 507 slides did not meet the AI quality requirements. Most were excluded because of blurred images, poor slide sealing, insufficient readable epithelial cells (fewer than 5,000 cells), or damaged slides. These findings highlight the importance of high-quality slide preparation for reliable AI analysis.

AI Demonstrated High Sensitivity for Cervical Lesion Detection

Among the 4,893 valid slides, the study identified:

• 4,119 negative cases
• 774 positive cases

Negative cases included NILM and microbial infections, including 5 trichomonas and 20 Candida infections.

Using manual diagnosis as the reference standard, the AI-assisted system achieved:

• 81% overall sensitivity
• 77% specificity

More importantly, the AI system reached 97% sensitivity when detecting LSIL and more severe lesions. However, its sensitivity for ASC-US remained relatively low.

These results indicate that AI performs particularly well in identifying clinically significant cervical lesions while reducing the likelihood of missing high-grade abnormalities.

False Positives Remain the Main Challenge

Although AI demonstrated excellent sensitivity, false positives were still observed.

The overall false-positive rate reached 23%. Most false-positive results occurred in ASC-US cases. In many instances, the system interpreted reactive epithelial cell changes as abnormal cells, leading to unnecessary positive classifications.

The AI system also classified 92 slides as LSIL. During manual review, pathologists found that glycogenated cells were frequently mistaken for koilocytes.

In addition, AI occasionally overestimated age-related atrophic squamous cells and densely clustered dark-staining cells as ASC-H or HSIL. These findings explain most of the false-positive results observed during the study.

AI Significantly Reduced Slide Reading Time

Reading efficiency showed one of the greatest advantages of AI-assisted screening.

According to laboratory quality control requirements, one pathologist should review no more than 100 slides per day.

Based on this standard, manually reviewing 5,400 slides would theoretically require approximately 4,320 working hours.

In actual practice, four pathologists completed the manual review in approximately 256 hours, or 32 working days.

By comparison, the KFBIO AI-assisted workflow—including whole slide scanning and automated image analysis—completed the same workload in only 162 hours (6.75 days).

Furthermore, unlike manual screening, the AI system can analyze slides continuously for 24 hours a day, further improving laboratory efficiency and throughput.

Key Findings

The clinical evaluation demonstrates several important advantages of AI-assisted cervical cytology screening:

• AI successfully analyzed more than 90% of submitted slides.
• The system achieved 97% sensitivity for detecting LSIL and more advanced lesions.
• Overall sensitivity reached 81%, while specificity reached 77%.
• AI reduced slide reading time by nearly 37% compared with manual review.

Continuous 24-hour analysis can significantly improve laboratory productivity and support large-scale cervical cancer screening programs.

Discussion

AI Shows Strong Potential for Large-Scale Cervical Cancer Screening

Cervical cancer is one of the few malignant diseases with a well-defined cause. More importantly, it is also one of the few cancers that can be effectively prevented and treated through early detection.

Over the years, cervical cancer screening has continued to evolve. New technologies, including DNA ploidy analysis, P16/Ki-67 dual staining, and methylation detection, have expanded the available screening options. However, because of cost and accessibility, cytological screening remains the most practical and cost-effective method in routine clinical practice.

AI Improves Diagnostic Efficiency While Maintaining High Sensitivity

This study evaluated 5,400 cervical cytology slides and compared AI-assisted screening with conventional manual slide interpretation.

The results demonstrated that the AI-assisted system achieved 77% specificity and 97% sensitivity for detecting LSIL and more advanced lesions. These findings suggest that AI can effectively support large-scale cervical cancer screening while maintaining high diagnostic performance.

In addition, AI can strengthen cytology quality control. Following the CNAS-15189 quality management requirements, laboratories can review all positive slides and a proportion of negative slides to ensure diagnostic consistency. As a result, AI can help laboratories improve both efficiency and quality assurance.

Current Limitations of AI-Assisted Cytology

Although AI demonstrated encouraging performance, several limitations remain.

High Requirements for Slide Quality

First, AI relies on high-quality digital slides.

In this study, 507 slides could not be analyzed because of poor image quality, sealing problems, insufficient readable epithelial cells, or damaged slides. These results indicate that standardized slide preparation and high-quality whole slide imaging remain essential for reliable AI-assisted diagnosis.

Lower Accuracy for ASC-US

Second, AI showed relatively lower accuracy when identifying ASC-US lesions.

Detailed analysis suggested that inflammatory reactive cells and slide impurities were occasionally classified as abnormal cells. Consequently, the false-positive rate increased. Similar diagnostic challenges also exist during manual slide interpretation because ASC-US remains one of the most subjective and least reproducible cytological categories.

Challenges with Atrophic Lesions

Third, AI sometimes overestimated age-related atrophic squamous cells and densely clustered cells as ASC-H or HSIL.

Several factors may contribute to these errors, including deeply stained nuclei, an increased nucleus-to-cytoplasm ratio, and morphological similarities between benign and abnormal cells.

These challenges are particularly relevant in population-based cervical cancer screening programs, where elderly women account for a large proportion of screened individuals. Therefore, future AI models should incorporate more representative datasets and continuously optimize their algorithms to improve diagnostic accuracy across diverse patient populations.

AI Is Becoming an Important Part of Digital Pathology

As artificial intelligence continues to develop, AI-assisted cytology is becoming an increasingly important component of digital pathology workflows.

Several international studies have reported comparable findings. For example, researchers in Japan demonstrated that computer-assisted cervical cytology achieved screening performance comparable to manual review while significantly reducing screening time. In addition, the U.S. Food and Drug Administration (FDA) has approved multiple commercial AI-assisted cervical cancer screening systems, highlighting the growing clinical acceptance of AI technologies worldwide.

Although AI has already been widely adopted in many developed countries, challenges remain in developing regions. Large populations, limited healthcare resources, and higher implementation costs continue to slow adoption. Therefore, developing cost-effective AI-assisted screening solutions will be essential for expanding access to high-quality cervical cancer screening in the future.

Future Outlook

AI will continue to improve as deep learning algorithms evolve and larger training datasets become available.

Rather than replacing pathologists, AI serves as an intelligent assistant that supports faster slide review, reduces repetitive workloads, and improves laboratory efficiency.

For large-scale cervical cancer screening programs and third-party pathology laboratories, combining AI-assisted analysis with expert pathologist review offers a practical approach to improving both diagnostic quality and operational efficiency.

Conclusion

AI Is Transforming Cervical Cytology Screening

AI-assisted cervical cytology offers an efficient and reliable approach to cervical cancer screening. By combining digital pathology with deep learning, AI can help pathologists identify abnormal cells faster while maintaining high diagnostic sensitivity.

In this study, the AI-assisted workflow demonstrated 97% sensitivity for detecting LSIL and more advanced lesions. It also significantly reduced slide reading time compared with conventional manual review. These findings highlight the value of AI in improving laboratory efficiency and supporting large-scale cervical cancer screening programs.

Although challenges remain, particularly in slide quality assessment and ASC-US interpretation, continued improvements in AI algorithms and larger clinical datasets are expected to enhance diagnostic performance.

Rather than replacing pathologists, AI serves as an intelligent clinical assistant. When combined with expert review and standardized quality control, AI-assisted digital pathology can improve diagnostic consistency, reduce workload, and optimize pathology workflows.

As digital pathology continues to expand worldwide, AI-assisted cervical cytology will play an increasingly important role in delivering faster, more accurate, and more accessible cervical cancer screening services.

Discover KFBIO‘s AI-Assisted Cervical Cytology Solution

KFBIO provides an end-to-end digital pathology solution for cervical cytology, including high-resolution whole slide imaging, AI-assisted analysis, pathology workflow management, and laboratory integration. Contact our team to learn how AI can improve your cervical cancer screening workflow.

FAQ

What is AI-assisted cervical cytology?

AI-assisted cervical cytology uses artificial intelligence and digital pathology technologies to automatically analyze cervical cytology slides. It helps pathologists identify suspicious abnormal cells more efficiently while improving workflow productivity.

How accurate is AI in cervical cancer screening?

In this study, the AI-assisted system achieved 97% sensitivity for detecting LSIL and more advanced lesions, with an overall sensitivity of 81% and a specificity of 77%.

Can AI replace pathologists?

No. AI is designed to assist pathologists rather than replace them. It performs automated pre-screening and highlights suspicious cells, while pathologists remain responsible for the final diagnosis.

What are the advantages of AI-assisted slide interpretation?

AI improves screening efficiency, shortens slide reading time, supports standardized quality control, and reduces repetitive workloads. These advantages make it particularly valuable for high-volume pathology laboratories.

Why is digital pathology important for cervical cancer screening?

Digital pathology enables whole slide imaging, remote collaboration, AI-assisted analysis, and standardized workflows. Together, these technologies improve diagnostic efficiency and support large-scale cervical cancer screening.