中文摘要

研究目的

本研究聚焦于当前人工智能（AI）技术在阴道镜检查中用于宫颈病变诊断所面临外部验证不足、功能较为单一以及在基层医疗机构中的应用泛化能力较弱等诸多问题，通过开展技术创新和临床验证工作，致力于提升AI在阴道镜检查全流程中的应用效能，为基层医疗提供智能化解决方案。首先，验证AI对宫颈上皮内瘤变2级及以上（CIN2+）检出效果及辅助医生诊断效能；其次，构建并优化宫颈转化区识别模型；最后，开展模型方法学的评价，测试模型在不同条件下的泛化能力，探索AI在拓展阴道镜检查维度中的可行性，为临床应用提供科学依据。

材料与方法

1. AI阴道镜诊断系统辅助检出宫颈病变的应用效果评价：基于公开独立的标准化阴道镜图像数据集187例，共859张阴道镜图像，评估本团队既往研发的AI诊断系统的泛化能力以及对阴道镜医生诊断效果的影响。分析不同年资经验的阴道镜医生在有无AI诊断系统辅助下的诊断效果，并采用多阅片者多病例研究设计，以组织病理检测结果为金标准，比较传统阴道镜诊断方式与 AI 诊断系统辅助阴道镜医生诊断对 CIN2 + 病例的检出效果。

2. 基于数字阴道镜图像的AI宫颈转化区识别模型的构建与验证：回顾性收集全国六家医院2278例女性的8335张阴道镜图像，按8:2划分为训练集和测试集用于构建AI转化区识别模型，并从训练集中随机抽取10%用于模型调优。基于测试集比较该模型与其他主流的神经网络对转化区类型的分类性能。

3. AI宫颈转化区识别模型外部验证研究与辅助阴道镜检查效果评价：回顾性收集全国4家医院1335例女性的6675张阴道镜图像。对AI转化区识别模型进行外部验证。随机选取210例病例，由4名医生在有无AI的辅助下独立判读，评估AI模型在识别转化区和阴道镜诊断中的辅助能力。

研究结果

1. AI阴道镜诊断系统辅助检出宫颈病变的应用效果评价：医生在阴道镜检查中对鳞柱交界可见性的判断的准确性为51.2%，对转化区类型判断的准确性为49.5%。在CIN2+为诊断阈值时，AI诊断系统与医生独立的检出敏感性均为84.2%。但AI 诊断系统的特异性较低（55.9% vs. 58.0%，p=0.049）。在AI诊断系统辅助后，医生检出CIN2+的敏感性提升至91.2%（p<0.001），尤其对低年资医生组的=提高更显著（AUC增加0.039, p=0.043）。此外，AI辅助医生对每例患者的平均活检数从2.48个降至2.02个，并显著提高低年资医生组的活检定位重合率（从0.80提升至0.86，p=0.031）。

2. 基于数字阴道镜图像的AI宫颈转化区识别模型的构建与验证：构建的AI转化区分类模型在测试集上实现了83.97%的分类准确率，其中TZ1、TZ2、TZ3的敏感性分别为84.74%、78.95%和87.87%，特异性分别为89.99%、91.98%和94.03%。对比试验显示，模型在分类准确率和精确率（83.93%）上显著优于ResNet50、VGG16和ViT等其他分类模型。此外，采用FastSAM的鳞柱交界识别模型在转化区分割任务中整体Dice系数为76.8%，平均精度为74.6%。通过Grad-CAM进行可解释性分析，模型生成的热力图与临床判断转化区的关键位置高度一致。

3. AI宫颈转化区识别模型外部验证研究与辅助阴道镜检查效果评价：AI模型在TZ1、TZ2和TZ3类型中的判别一致性分别为77.3%、81.1%和80.3%。在TZ1中，AI判别特异性为94.2%，主要误判为TZ2。在TZ2病例中，AI特异性为83.3%，与TZ3存在混淆。TZ3判别中，AI特异性为90.7%，但误判为TZ2的比例较高。AI模型在所有转化区类型中的敏感性均显著高于不同年资医生。低年资和高年资医生联合AI检出CIN2+的敏感性均高于独立诊断，在TZ3病例中，低年资医生联合AI转化区诊断的敏感性为87.0%，独立诊断为78.3%；而高年资医生联合AI对CIN2+的敏感性大幅高于其独立诊断。

研究结论

1. AI阴道镜诊断系统辅助检出宫颈病变的应用效果评价：基层医生的总体阴道镜检查能力依然较差。AI诊断系统显著提升了阴道镜医生检出CIN2+的准确性和效率，尤其能够均衡不同经验医生的检出水平，降低漏诊风险。AI技术可作为优化宫颈癌筛查策略的有效工具，特别适用于低卫生资源地区。

2. 基于数字阴道镜图像的AI宫颈转化区识别模型的构建与验证：基于AI算法构建的转化区识别模型在多中心测试集中展现出优异的分类和识别性能，能够准确识别不同形态和位置的转化区，并在多样化临床数据中表现出良好的预测能力。模型结合转化区分型与新鳞柱交界位置，显著提升了阴道镜医生识别转化区的效率和准确性，为活检指导和精准治疗提供了可靠支持。

AI宫颈转化区识别模型外部验证研究与辅助阴道镜检查效果评价：本研究构建的基于深度学习的AI模型在转化区分类任务中表现优异，准确性高于我国低卫生资源地区的临床医生，展现了其临床应用潜力。模型具备轻量级和低算力需求，适配于低资源地区的阴道镜辅助检查。未来需通过前瞻性临床试验和模型优化，进一步验证其在医疗环境中的真实适用性，并提升对分类模糊任务的性能。

Abstract

Objectives

This study addresses limitations of current AI technologies in colposcopic examination, such as insufficient external validation, limited functionality, and poor generalizability in primary healthcare settings. Through technological innovation and clinical evaluation, the study was to enhance the integration of AI into the colposcopy workflow, offering intelligent solutions for primary care and improving the accuracy and efficiency of early diagnosis. The performance of the AI system in detecting CIN2+ lesions and its effectiveness in assisting colposcopists were systematically evaluated. Furthermore, an AI-based cervical transformation zone recognition model was developed and optimized by integrating multidimensional features and enabling visualization of the squamocolumnar junction. External validation was conducted to assess the model’s robustness and generalizability across various clinical conditions and digital images from different colposcopic devices, supporting the feasibility of AI in enhancing colposcopic capabilities and providing scientific evidence for clinical implementation.

Methods

1. Evaluation of the effect of AI-assisted colposcopy diagnostic systems in detecting CIN2+ lesions: A total of 859 colposcopic images from 187 cases within a publicly available, independent, and standardized colposcopy image dataset were utilized to evaluate colposcopists' varied examination capabilities and the generalization ability of the AI diagnostic system previously developed, and whether colposcopists with varying levels of experience show improved diagnostic performance with AI assistance. A multireader, multicase study design was employed to compare the detection performance of CIN2+ cases between traditional colposcopic diagnosis and AI-assisted colposcopic diagnosis. The diagnostic performance of colposcopists with and without AI assistance were evaluated by final histopathological results as the gold standard.

2. Development and validation of an AI-based cervical transformation Zone identification model: This study retrospectively collected digital colposcopy images and medical records from 2,278 female patients across six hospitals in China, including cytology, HPV infection status, colposcopy findings, and pathological diagnosis results. All images were acquired using standardized high-definition digital video colposcopes and underwent rigorous quality control by experts in accordance with IFCPC guidelines. A total of 8,335 images were included and partitioned into a training set and a test set with an 8:2 ratio for model training. 10% of the training set was randomly allocated as a validation set for model optimization. The classification performance of the developed model in classifying transformation zone types was compared against that of state-of-the-art classification neural networks using the internal test set data.

3. External validation of AI-based cervical transformation zone identification model and its effect on assisting colposcopy: This retrospective study collected colposcopy images from patients referred for colposcopy due to cervical cancer screening and undergoing cervical lesion treatment from four hospitals in China. A total of 1,335 cases with 6,675 colposcopy images were included. The validation of the AI transformation zone identification model was performed against the gold standard determined by an expert panel based on IFCPC guidelines. Performance metrics, including sensitivity, specificity, accuracy, and AUC, were calculated. Subsequently, 210 cases were randomly selected from the validation dataset and four colposcopists with varying experience levels were stratified into AI-assisted and non-AI-assisted groups for independent interpretation. A comparative analysis of interpretation results between the two groups was conducted to assess the feasibility and effectiveness of the AI model in supporting clinical decision-making.

Results

1. Evaluation of the effect of AI-assisted colposcopy diagnostic systems in detecting CIN2+ lesions: The accuracy of colposcopists in assessing the visibility of the squamocolumnar junction during colposcopic examination was 51.2%, while the accuracy for determining the transformation zone type was 49.5%. When using CIN2+ as the diagnostic threshold, the sensitivity of detection by both the AI-assisted diagnostic system and independent colposcopists was 84.2%. However, the AI system demonstrated lower specificity compared to physicians (55.9% vs. 58.1%, p=0.049). After incorporating AI assistance, the sensitivity of colposcopists in detecting CIN2+ increased significantly to 91.2% (p<0.001), with a particularly notable improvement observed among less-experienced doctors (increase in AUC by 0.039, p=0.043). Additionally, AI assistance optimized biopsy efficiency, reducing the average number of biopsies per case from 2.48 to 2.02, and significantly enhanced biopsy site localization accuracy in the junior colposcopist group, with the median concordance rate increasing from 0.80 to 0.86 (p=0.031).

2. Development and validation of an AI-based cervical transformation Zone identification model: The proposed transformation zone classification model achieved a test set accuracy of 83.97%. The model demonstrated sensitivities of 84.74%, 78.95%, and 87.87% for TZ1, TZ2, and TZ3, respectively, and specificities of 89.99%, 91.98%, and 94.03%. Comparative analysis revealed that this model outperformed ResNet50, VGG16, and ViT in terms of both accuracy (83.97%) and precision (83.93%). Additionally, the FastSAM was performed exhibited a Dice score of 76.8% and a mean average precision of 74.6% in transformation zone segmentation. Grad-CAM analysis demonstrated expert-identified features for transformation zone determination, supporting the model's decision-making interpretability.

3. External validation of AI-based cervical transformation zone identification model and its effect on assisting colposcopy: The model demonstrated consistency of 77.3%, 81.1%, and 80.3% in identifying TZ1, TZ2, and TZ3, respectively. For TZ1 cases, the model exhibited high specificity (94.2%) , with most misclassifications occurring as TZ2. In TZ2 cases, the model achieved a specificity of 83.3%, though some confusion with TZ3 was observed. Similarly, for TZ3 cases, the model demonstrated a specificity of 90.7%, with a notable proportion of misclassifications as TZ2. The AI model consistently outperformed colposcopists with varying levels of experience in sensitivity across all TZ types. The diagnostic sensitivity for CIN2+ lesions was higher when both junior and senior colposcopists were assisted by the AI cervical TZ identification model compared to when they performed independent diagnoses. In TZ3 cases, the sensitivity of junior colposcopists in diagnosing CIN2+ lesions with AI assistance based on transformation zone classification reached 87.0%, compared to 78.3% with independent assessment. For senior colposcopists, AI-assisted diagnosis achieved a highest sensitivity for detecting CIN2+ lesions.

Conclusions:

1. Evaluation of the effect of AI-assisted colposcopy diagnostic systems in detecting CIN2+ lesions: The AI diagnostic system significantly improved colposcopists' accuracy and efficiency in detecting CIN2+, particularly by balancing detection performance across different experience levels and reducing the risk of missed diagnoses. However, the overall colposcopy examination performance of primary care colposcopist remains suboptimal. AI technology can serve as an effective tool for optimizing cervical cancer screening strategies, especially in low-resource settings.

2. Development and validation of an AI-based cervical transformation Zone identification model: The AI algorithm-based transformation zone identification model exhibited superior classification and recognition performance in a multi-center test set. The model accurately identified transformation zones of diverse morphologies and locations, demonstrating robust predictive capability across a heterogeneous clinical dataset. The integration of transformation zone subtypes and precise squamocolumnar junction localization significantly enhanced the efficiency and accuracy of transformation zone identification by colposcopists, providing reliable support for biopsy guidance and precision treatment.

3. External validation of AI-based cervical transformation zone identification model and its effect on assisting colposcopy: The deep learning-based AI model that demonstrated excellent performance in TZ classification, achieving higher accuracy than clinical colposcopists in low-resource regions of China, highlighting its potential for clinical application. The model is lightweight and requires low computational requirement, making it suitable for AI-assisted colposcopic diagnosis in resource-limited settings. Future work should focus on prospective clinical trials and model optimization to further validate its real-world applicability in medical environments and enhance its performance in ambiguous classification tasks.