Cutting-edge progress of artificial intelligence in cervical cytology

Cervical cancer ranks as the fourth most common cancer among women, with the incidence and mortality rates of cervical cancer remaining particularly high in low- and middle-income countries.1 This may be related to factors such as ineffective vaccination programmes and a shortage of skilled, professional cytology personnel.2 3 In many countries, including China, cytopathology as a sub-specialty has not been completely established, and the cytotechnologist career is virtually non-existent. With the development of artificial intelligence (AI) technology, new technological approaches have emerged to improve Pap test screening, which is a critical component of Pap test review, but can be repetitive and labour-intensive. Currently, cervical cytology AI screening systems are divided into two major categories: slide-based AI systems and whole slide imaging (WSI)-based AI systems.

The US Food and Drug Administration has successively approved at least three digital pathology systems for clinical use: the PAPNET system, FocalPoint GS system and Hologic ThinPrep Imaging System (TIS), all of which fall under the category of slide-based AI systems.4 The PAPNET system employs neural network technology to identify abnormal epithelial cells in Pap smears, effectively distinguishing between normal and abnormal cells while significantly reducing false negative rates, and is primarily applied in quality control processes.5 The BD FocalPoint GS system has demonstrated a sensitivity equal to manual screening in diagnosing low-grade squamous intraepithelial lesions and above (LSIL+) (88% for BD FocalPoint GS vs 89.7% for manual screening, p>0.05) and high-grade squamous intraepithelial lesions and above (HSIL+) (83.8% for BD FocalPoint GS vs 83.3% for manual screening, p>0.05).6 Additionally, the TIS-assisted screening system nearly doubles slide reading efficiency while maintaining sensitivity equal to or exceeding that of manual primary screening, with no adverse impact on specificity.7

WSI-based AI systems have achieved a transformation from traditional microscopic operations to digital screening, enabling pathologists to directly observe key images selected by the system on computer screens. Deep learning-based AI systems can directly analyse digitised Pap slides, automatically identifying cells relevant to diagnosis through AI technology. These systems can also estimate squamous cell counts and detect glandular cells and infections, thereby providing diagnostic fields for the cytopathologists’ final review. Despite these updates, sample collection and slide preparation procedures remain unchanged.

Currently, the market encompasses multiple WSI-based cervical cytology screening AI platforms developed across various countries, including CytoProcessor (DATEXIM, Caen, France), BestCyte Cell Sorter Imaging System (CellSolutions, Greensboro, USA), Genius Digital Diagnostics System (Hologic, Marlborough, USA), Techcyte SureView Cervical Cytology System (Techcyte, Orem, USA) and CytoSiA Pro (OptraSCAN, Pune, Maharashtra, India).4 Additionally, China has emerged as a significant contributor to this field with several domestically developed platforms, namely Landing Med, KFBIO, AICCS, CIAS, AICyte, AIATBS and CITL-AI.4 8 This proliferation of AI-based screening platforms demonstrates the global recognition of AI’s potential in enhancing cervical cancer screening efficiency and accuracy. Through automated slide scanning, AI platforms can identify ‘Objects of Interest (OOIs)’ for cytologists to examine and render a final diagnosis.

Currently, AI platforms operate in three modes: the first is the assistance mode, where systems like CytoProcessor,9 BestCyte Cell Sorter Imaging System10 and Hologic Genius Digital Diagnostics System (HGDDS)11 serve only as assistant tools without directly providing diagnostic results. For example, HGDDS can provide 30–60 OOIs, and cytologists can make the diagnosis based on the AI selected OOIs. HGDDS could maintain sensitivity, specificity and accuracy for the detection of cervical dysplasia, but significantly reduce diagnostic time compared with TIS.11 One study demonstrated that higher sensitivity was observed across the four tested commonly identified micro-organisms (Actinomyces, Candida, Trichomonas and Herpes virus) on HGDDS.12 This model can improve work efficiency and easily be accepted in countries with highly trained cytology personnel.

The second mode is that the AI system automatically makes the diagnosis of negativity and positivity for each slide, in addition to providing the user with OOIs. The negative cut-off value will depend on the modes, such as 37.5% in CITL-AI13 and 50% in AICyte.14–16 An analysis of 163 848 cases from four institutes in China by using 50% cut-off value demonstrated that the AICyte alone could be an independent screening tool for pathological practices that do not have cytotechnologists or where the workload is heavy.16 This may be particularly applicable in countries lacking cytotechnologists and facing heavy cervical cancer screening workloads, such as China and India.

The third mode is that an AI platform can provide specific diagnoses for each case according to the TBS classification system, including negative, atypical squamous cells of uncertain significance, LSIL, atypical squamous cells—cannot rule out high-grade lesion and HSIL. Representative systems using this mode include AICCS,17 AIATBS18 and CIAS.19 We think an AI system cannot replace cytologists, and this mode is still far from ready for clinical diagnostic application. Regardless of mode, AI platforms significantly improve diagnostic efficiency compared with traditional manual slide reading, with average processing time per slide ranging from 22.23 to 180 s.11 14 18–20 Additionally, AI platforms generally demonstrate higher diagnostic performance; a statistical study showed different AI platforms showed sensitivity ranges of 73%–95.9%, specificity ranges of 49.7%–99.4% and accuracy ranges of 85.7%–88.1% for detecting high grade cervical lesions when ASC-US and above abnormal Pap were considered.8 The studies from European and American patient populations (HGDDS, CytoProcessor, BestCyte Cell Sorter Imaging System) indicated the average sensitivity for detecting high-grade cervical lesions was 93.3% for AI-assisted and 90.8% for manual Pap test review.8

Despite the significant advantages demonstrated by AI technology in cervical cytology screening, its clinical deployment still faces multiple challenges. First, AI platforms require substantial costs for equipment procurement and infrastructure construction, posing barriers for resource-limited medical institutions. Second, AI system performance is highly dependent on training data quality; differences in populations and regions, as well as slide preparation techniques, may lead to insufficient model generalisability. Third, different AI platforms exhibit significant diagnostic variability in detecting high-grade lesions, with sensitivity ranging from 73% to 95.9% and specificity from 49.7% to 99.4%. Additionally, the approval process for AI medical devices is complex, lacking unified diagnostic standards and quality control guidelines, which affects the standardised promotion of the technology. Furthermore, recent studies have been predominantly retrospective; there is a lack of prospective studies specifically designed for AI platforms with long-term follow-up data to validate real-world diagnostic performance. To address these challenges, it is necessary to establish the standardised approval processes and industry regulations, develop quality control standards, conduct multicentre prospective studies and strengthen technical training and education for medical personnel.

In the future, the development of AI platforms should strike a balance between technical optimisation and clinical applicability. On the one hand, through continuous optimisation of deep learning algorithms and expansion of training datasets, AI platforms are expected to further improve their negative exclusion rates. This will significantly reduce the workload of cytologists and enhance screening efficiency. On the other hand, it is even more important to consider human-machine collaboration models, where AI-assisted screening and preliminary classification are combined with final review and confirmation by cytologists, thereby leveraging AI’s efficiency advantages while ensuring diagnostic accuracy and reliability. This would achieve an organic integration of technological advancement with clinical practice and provide practical and feasible solutions to address issues of uneven medical resource distribution and shortage of specialised personnel.