Cervical cancer is a major global health concern, ranking among the most common cancers and leading causes of cancer-related deaths. Early detection through screening is crucial for prevention and improved prognosis. Current screening methods include cervical cytology, HPV testing, and DNA ploidy testing, with cervical cytology being widely used due to its simplicity and cost-effectiveness. However, a significant shortage of cytopathologists worldwide, particularly in countries like China, leads to a high false-negative rate. This shortage highlights the urgent need for auxiliary tools to improve the accuracy and efficiency of cervical cancer screening. Artificial intelligence (AI), particularly deep learning techniques, has shown promise in medical image analysis and diagnosis across various cancers, including diabetic retinopathy, lung cancer, and breast cancer. The application of AI to cervical cytology offers potential advantages such as increased consistency, improved alignment with biopsy outcomes, higher sensitivity, and reduced misdiagnosis. Previous AI-based approaches for cervical cytology analysis have either faced limitations in broad applicability due to increased complexity or focused solely on specific cell types. This study aimed to develop and validate an AI-powered cervical cancer screening (AICCS) system to assist in classifying cervical cytology cases and cervical cancer by analyzing whole-slide images (WSIs). The system was validated using multicenter, retrospective, prospective, and randomized observational trial data.

The literature reveals a critical need for improved cervical cancer screening due to the global burden of the disease and the limitations of current methods, especially the shortage of trained cytopathologists. Several studies have explored the use of AI in cancer diagnosis, demonstrating its potential to improve accuracy and efficiency in various settings. The application of AI to cervical cytology has shown promise in previous research, although limitations exist in terms of generalizability and focus on specific cell types. This study builds upon this existing work by developing a comprehensive AI system validated across multiple datasets and methodologies, addressing the limitations of previous approaches. The review highlights the growing interest in using AI to address the challenges in cervical cancer screening, positioning this study as a significant contribution to this field.

The study employed a three-phase approach: proof-of-concept (POC), validation, and a randomized observational trial. The AICCS system comprises two main AI models: a patch-level cell detection model and a WSI-level classification model. The patch-level model, using RetinaNet, identifies abnormal cells within smaller image patches. The WSI-level model, employing a random forest algorithm, classifies the entire slide based on the patch-level results. Data from 16,056 participants were used, including retrospective and prospective datasets from multiple institutions, along with a randomized observational trial. The datasets were carefully curated and balanced to address class imbalance issues in the training data. The AICCS system was trained and validated using various deep learning techniques and optimized using different algorithms. The performance of the AICCS was evaluated using metrics such as AUC, sensitivity, specificity, and accuracy across various datasets and subgroups. The study rigorously adhered to guidelines such as STRAD, CONSORT-AI extension, and the MI-CLAIM checklist. A key aspect of the methodology involved the quality control measures for WSI acquisition and preprocessing, including an AI-assisted approach for identifying and addressing scanning quality issues. Detailed annotation protocols were followed to ensure the accuracy and consistency of the training data. The study also addressed the problem of class imbalance in the data through data augmentation techniques and careful dataset construction.

The AICCS system consistently demonstrated high accuracy in predicting cervical cytology grades across different datasets. In the prospective assessment, it achieved an area under the curve (AUC) of 0.947, a sensitivity of 0.946, a specificity of 0.890, and an accuracy of 0.892. The randomized observational trial revealed significantly higher specificity, accuracy, and sensitivity (a 13.3% increase) for AICCS-assisted cytopathologists compared to cytopathologists alone. Subgroup analysis showed high AUC values across all subgroups, with improvements in AUC corresponding to worsening cytological grades. The AICCS system showed comparable performance in specificity and accuracy across all internal and external validation datasets, maintaining values above 0.810 for both metrics. The AICCS also exhibited high negative predictive values (NPV) for the validation datasets. Comparison with cytopathologists using histopathological diagnoses as the gold standard showed no significant difference in detecting abnormal cytology grades between AICCS alone and cytopathologists. The AICCS significantly reduced the diagnostic process time, completing analysis in under 120 seconds compared to approximately 180 seconds for manual reading. The study provides a web-based platform for clinical application of the system.

The findings demonstrate the significant potential of AI-assisted cervical cytology screening for improving accuracy and efficiency in cervical cancer detection. The high accuracy and performance of the AICCS system across various datasets and subgroups validates its robustness and generalizability. The improved performance of AICCS-assisted cytopathologists compared to cytopathologists alone highlights the system's ability to augment human expertise and reduce the impact of inter-observer variability. The substantial reduction in diagnostic time offers a significant advantage in clinical practice. The successful implementation of the cloud-based platform facilitates broader access to this technology, potentially addressing the challenges of cytopathologist shortages in resource-limited settings. The results contribute substantially to the field of AI in healthcare and provide a strong foundation for further development and implementation of AI-based solutions for cervical cancer screening.

This study presents a novel AI-assisted cervical cytology screening system (AICCS) demonstrating high accuracy and efficiency in detecting cervical cancer. The AICCS outperforms cytopathologists alone and significantly enhances the performance of cytopathologists when used as an assistive tool. This technology holds the promise of improving cervical cancer screening, especially in regions with limited access to specialists. Future research should focus on larger-scale clinical trials and integrating the AICCS into routine clinical practice to assess its long-term impact and cost-effectiveness.

Limitations include the potential for inaccurate sample representation due to limited operator experience and variability in specimen preparation across medical centers. These factors could affect the AI system's accuracy and highlight the need for comprehensive evaluation and monitoring. The study also notes the importance of considering privacy and security concerns associated with AI-based healthcare systems.