This article is from Chinese Clinical Oncology

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

Cervical cancer is one of the leading causes of cancer-related deaths among women in both China and worldwide. It ranks as the fourth most common malignant tumor in women globally. In 2021, there were 109,741 newly diagnosed cases of cervical cancer in China, with 59,060 deaths, placing cervical cancer as the eighth leading cause of cancer-related deaths in women.

Adhering to the principle of prevention as the primary approach and combining prevention with treatment, increasing the coverage of cervical cancer screening, innovating screening models, and improving the quality and efficiency of screening can result in an early detection rate of over 90% for cervical cancer. Currently, there is a significant shortage of pathologists in China, leading to high case review pressures. When calculated at a rate of 100 slides per pathologist per day, it falls far short of meeting clinical demands. Moreover, there is an imbalance in the distribution of medical resources across regions, with significant differences in the understanding of cytology quality control. Due to factors such as fatigue, skill levels, and subjective interpretation, the detection sensitivity is only around 65%, posing a risk of both missed diagnoses and misdiagnoses, thereby compromising the quality of cytological diagnoses.

Therefore, improving screening sensitivity, specificity, and efficiency is an urgent issue to address. With the advent of the era of computer science and data collection, artificial intelligence (AI) technology has gradually expanded and integrated into various medical fields. In particular, in the field of cytology, AI-assisted slide interpretation combines human thinking with the capabilities of artificial intelligence, creating an excellent cytological screening method. This significantly reduces the screening time for pathologists and selects individual abnormal cells, thereby improving the efficiency of information processing in cytological screening. This study compares the detection of cervical precancerous lesions in AI-assisted slide interpretation systems with conventional microscopic slide interpretations, examining their sensitivity and specificity to validate their application value in cervical cytology.

1. Clinical Data Collection: A total of 5,400 diagnosed and manually reviewed ThinPrep Cytologic Test (TCT) slides, with satisfactory quality, archived by Huikang Medical Laboratory from January 2022 to August 2022, were selected as the control group for the study. Additionally, 4,893 slides identified as valid by AI were used as the AI group.

2. Cytological Diagnostic Criteria: The 2014 TBS [11] descriptive diagnostic criteria were employed, categorizing atypical squamous cells of undetermined significance (ASC-US) and above as positive, and other categories as negative. Specific diagnostic terms included: (1) No intraepithelial lesion or malignancy (NILM); (2) ASC-US; (3) Atypical squamous cells, cannot exclude high-grade squamous intraepithelial lesion (ASC-H); (4) Low-grade squamous intraepithelial lesion (LSIL); (5) High-grade squamous intraepithelial lesion (HSIL); (6) Squamous cell carcinoma (SCC); (7) Atypical glandular cells – not otherwise specified (AGC-NOS); (8) Atypical glandular cells – favor neoplasm (AGC-FN); (9) Adenocarcinoma in situ (AIS); (10) Adenocarcinoma. Due to the rarity of glandular epithelial lesions in actual work, only 5 cases of AGC cytology reports were issued by the laboratory from January to August 2022, so glandular epithelial lesions were not included in this study.

3. Research Methods: Automated sedimentation-based slide preparation was used for cytological examination. Manual slide interpretation was conducted by two attending physicians and two associate chief physicians. Scanning of slides was performed using the KF-PRO-400 scanner from KFBIO, and AI-assisted cervical cancer diagosis software from KFBIO.

2. Results:

1. Evaluation of AI Slide Quality: A total of 5,400 slides were scanned and identified, with 4,893 slides recognized as valid by AI. These included 4,094 NILM cases, 20 cases of fungal infection, 5 cases of trichomonas infection, 280 cases of ASC-US, 267 cases of LSIL, 67 cases of ASC-H, 158 cases of HSIL, and 2 cases of SCC. A total of 507 slides were deemed unfit for analysis due to reasons such as blurry images caused by focal or sealing issues, insufficient readable epithelial cell quantity (less than the AI-set standard of 5,000 cells), and damaged slides.

2. Sensitivity and Specificity Analysis of AI-Assisted Slide Interpretation: In this study, 4,119 negative slides and 774 positive slides were identified. Negative slides included NILM and microbial infections, with 5 cases of trichomonas infection (Figure 1A) and 20 cases of candida infection (Figure 1B) among the 4,119 negative slides. Using the control group as a standard, the sensitivity of the AI group was 81%, and the specificity was 77%. AI achieved the highest sensitivity of 97% in identifying LSIL and above lesions, with the lowest sensitivity observed for ASC-US. In negative slides, AI often misinterpreted mucus threads or amorphous rod-shaped structures as candida or trichomonas, though limited by the scarcity of comparable data for microbial infections.

3. False Positives and False Negatives Analysis in the AI Group: The false positive rate in the AI group was 23%, primarily attributed to misidentifying ASC-US cases, with 724 cases of reactive changes misinterpreted as pathological cells (Figure 1C), contributing to the increased false positive rate. Additionally, 92 cases were interpreted by AI as LSIL. During the manual review, it was observed that most cases involved AI misinterpreting glycogenated cells as hollow cells (Figure 1D). In 76 cases of ASC-H and 44 cases of HSIL lesions, the majority were found to be basal cells undergoing age-related atrophy (Figure 1E) or compressed dark-staining cell clusters (Figure 1F).

A: Trichomonas; B: Candida; C: AI interprets reactive changes in epithelial cells as ASC-US; D: AI interprets glycogenated cells as low-grade lesions (LSIL); E: AI interprets atrophic squamous epithelium as high-grade lesions (HSIL); F: AI interprets crowded cell clusters as high-grade lesions (HSIL).

2.4 Slide Reading Time Comparison: According to quality control requirements, each person’s daily slide reading should not exceed 100 slides. The theoretical total time for manual slide reading was calculated as 4,320 hours. In actual practice, four pathologists collectively reviewed 5,400 slides, taking approximately 256 hours (32 days) in total. The AI system, including slide scanning and analysis, required only 162 hours (6.75 days). AI significantly shortened the slide reading time, and it can analyze slides continuously for 24 hours.

3. Discussion

Cervical cancer is currently the only malignant tumor with a relatively clear cause and the only one that can be cured and prevented. Cervical cancer screening has been conducted for many years and has given rise to various new technologies such as DNA ploidy analysis, P16/Ki-67 dual staining, and methylation detection. However, due to economic costs and universality issues, cytological screening remains the most economically effective method at the current stage.

This study analyzed 5,400 cytological smears, comparing the slide reading time, sensitivity, and specificity between the AI group and the control group. The results showed that the specificity of the AI group reached 77%, and the sensitivity for LSIL and above lesions could reach 97%. This indicates that, to some extent, AI assistance can be relied upon for large-scale cervical cancer screening, and effective cytological quality control (following CNAS-15189 requirements, reviewing all positive slides and 10% of negative slides) can ensure diagnostic quality. However, compared to traditional slide reading, AI still has certain limitations. Firstly, AI has higher requirements for slide quality; in this study, 507 slides could not be interpreted due to quality issues, highlighting the high recognition rate of AI for high-standard, high-quality slides. Secondly, AI has a lower accuracy in interpreting ASC-US lesions. The analysis indicates that AI misinterprets some inflammatory reactive cells or impurities as pathological cells, leading to an increased false positive rate. Such lesions also exhibit characteristics of strong subjectivity and poor repeatability in manual slide interpretation. In practical work, AI often plays the role of a “garbage bin.” Thirdly, AI shows significant misjudgment in age-related atrophic lesions and crowded cell clusters. The reasons for this may be deep staining of atrophic cell nuclei, a nucleus-to-cytoplasm ratio imbalance, and morphological overlap with highly pathological cells. However, in government-supported cervical cancer screening, which often takes place in economically challenged areas with a higher proportion of elderly individuals, AI needs further algorithm refinement and increased data matching to improve interpretative accuracy. As AI technology continues to evolve, cytological screening systems with intelligent assistance in slide interpretation are becoming increasingly popular. The successful application of automated screening in intelligent-assisted slide interpretation will provide new ideas and methods for the development and application of pathological intelligent software. Moreover, this technology has already been widely applied in developed countries. Japanese scientists, such as Tanaka et al. [7], compared manual slide interpretation with computer imaging screening for over four thousand cervical cytological smears, and the results showed that the cytological screening capability of this system was comparable to manual screening, with a significantly shorter screening time. Therefore, some medical laboratories in developed countries have purchased this system and incorporated it as part of routine cervical cytological screening. The U.S. Food and Drug Administration (FDA) has approved several commercialized cervical cancer screening systems, such as Becton Dickinson’s FocalPoint GS Imaging System [8], making AI widely used in the field of cytological screening. However, in developing countries like China, where the population base is large and the maintenance and usage costs of these auxiliary systems are high, widespread application is challenging. Therefore, it is hoped that shortly, AI screening systems with higher cost-effectiveness can be widely applied at the grassroots level.

In summary, while continuously simulating human experience through deep learning, AI can gradually improve computer algorithm models. This improvement leads to an enhancement in cell diagnostic effectiveness. As a form of computerized big data screening method, AI significantly shortens the slide reading time for pathologists, reduces their workload, and thereby improves the efficiency of pathologists in cytological screening. In practical work, making full use of AI technology while maintaining pathologists as the leaders, and conducting effective cytological quality control, especially for large-scale screening work and third-party pathological diagnosis centers, can significantly enhance work efficiency.