This retrospective study included 3616 women who underwent colposcopy examinations across 6 multicenter hospitals in China between January 2019 and October 2021. These hospitals were the Gansu Maternity and Child Healthcare Hospital, Second Hospital of Shanxi Medical University, Shenzhen Maternity and Child Healthcare Hospital, Jiangxi Maternity and Child Healthcare Hospital, The Affiliated Hospital of Qingdao University, and Chengdu Women’s and Children’s Central Hospital. The digital clinical records, including cytology, human papillomavirus infection status, colposcopy findings, and pathological results, were collected. All colposcopy images were captured using digital high-definition video colposcopes (Zonsun Healthcare Co Ltd, Edan Instruments). Colposcopy findings, including adequacy, SCJ visibility, TZ determination, and provisional diagnosis, were qualified by colposcopy experts from tertiary hospitals. The “ground truth” for TZ classification and SCJ visibility in this study was determined by an expert panel following IFCPC guidelines.

The inclusion criteria were as follows: (1) women with complete colposcopy findings and pathological diagnosis, (2) ages between 24 and 65 years, and (3) each record containing at least 5 satisfactory colposcopic images before and after acetic acid staining. The exclusion criteria were (1) all saline solution images (preacetic), since the TZ typically appears after acetic acid staining; (2) poor-quality images, such as overexposed images or those where the cervix was obscured or there was bleeding after the biopsy; (3) inadequacy colposcopic examination; and (4) records with missing TZ types or SCJ visibility. For external validation, an independent dataset was derived from The Affiliated Hospital of Qingdao University and Chengdu Women’s and Children’s Central Hospital. The training dataset was divided into training and test sets in an 8:2 ratio, with 10% of the training set used for validation during model tuning.

The colposcopy procedures adhered to the standard guidelines for Colposcopy and Treatment of Cervical Precancer [6]. Time-series images were captured for each case, including 1 original cervix image (saline solution) and at least 4 acetic acid–stained cervix images. The classification of TZ followed guideline criteria. TZ was classified as type 1 when it was entirely located on the ectocervix. In type 2, TZ was partially located within the endocervical canal, but its upper limit could be visualized using auxiliary instruments. Type 3 TZ lies partly or entirely inside the endocervical canal, with its upper limit being partially or completely invisible, even with the aid of auxiliary instruments.

The colposcopy images were captured in high resolution. However, they contained irrelevant objects, such as the endocervical speculum, cotton swabs, or large regions of the vaginal wall, which could interfere with the classification model’s ability to extract critical features of the cervix. To address this issue, we used the YOLOv5 network to detect the region of interest. Its single forward pass design enables real-time object detection with high efficiency and precision [16,17]. The integrated cervical region was automatically segmented and examined by a specialist (Figure 1). It divides an image into an S×S grid cell, predicting the bounding box locations and their associated categories. All segmented colposcopy images were resized to 224×224 pixels to align with the input specifications of the model.

Initially, colposcopy images were input into a detection model to determine the cervical region, which was then preprocessed to enhance feature extraction. These data were then applied to a classification model to determine the types of TZ present (Figure 2A). In the second part of the method, the original images were annotated with new SCJ prompts to guide a general segmentation model in inferring the potential location of the TZ (Figure 2B).

We proposed a variant of the MobileNetV3 architecture for classifying TZ, a lightweight convolutional neural network (CNN) specifically designed for efficient operation on portable devices, which was particularly suitable for deployment in resource-limited settings. The overall model structure is presented in Figure 3. Our model retained the depthwise separable convolutions and the HardSwish (H-swish) activation function of MobileNetV3, which reduced computational demand and parameters without compromising accuracy. Multiscale convolution modules have been incorporated into the model to effectively extract features at multiple focal points (Figure 3A), from subtle acetowhite changes in the columnar epithelium to varying TZ-type scopes in the cervix. In addition, we introduced a spatial pyramid pooling module to address features at multiple scales while preserving spatial information in the input images to ensure the richness of feature representation (Figure 3B). The Squeeze-and-Excitation module is a lightweight attention mechanism designed to automatically prioritize the most diagnostically relevant visual patterns within the MobileNetV3 architecture (Figure 3C). In our model, the Squeeze-and-Excitation module was embedded to adjust global feature weights (Figure 3D), enhancing the overall effectiveness of feature extraction. It increased the sensitivity to important features, thereby improving classification performance [18].

The FastSAM model was used to segment the TZ from the entire cervix in colposcopy images (Figure 2B). FastSAM was fine-tuned for the specific dataset using a YOLO-based framework [19]. Colposcopic images were resized to 640×640 pixels, and the new SCJ was annotated using LabelMe (MIT CSAIL) [20]. By selectively freezing some of the backbone layers of the model, the deeper layers were fine-tuned to improve adaptation to the specific task, without compromising the extraction of general features. The model’s performance was further enhanced by using coarse masks as spatial guides, which allowed the model to focus its attention on the most relevant image regions. The model was trained to delineate the new SCJ, thereby identifying possible TZ areas by this critical boundary. TZ is a dynamic region defined by the area between the original SCJ and the new SCJ. Typically, the original SCJ is considered a virtual line, and identifying the new SCJ is crucial for determining the location of the TZ (Figure 4).

The study was performed on a Windows 11 operating system with an NVIDIA GeForce RTX 3080 8GB graphics card. The models were implemented using Python (version 3.6.13; Python Software Foundation) and the PyTorch (Meta AI) Library (version 1.7.1).

Experiment performance of the cervical TZ classification model was evaluated using accuracy, precision, recall, F₁-score, and average precision, as defined in Equations 1-6.

The performance of the classification model was evaluated by comparing it with selected state-of-the-art (SOTA) deep-learning models using metrics such as accuracy, precision, recall, and the area under the precision recall curve, represented by average precision. To evaluate the classification of TZ types, sensitivity, specificity, positive predictive value, and negative predictive value (NPV) were calculated, along with their corresponding 95% CIs. The evaluation metric was defined as the agreement with the expert-provided TZ classification. The demographic characteristics of the study participants were summarized using means and SDs for continuous variables and percentages for categorical variables. A P value of less than .05 (two-sided) was considered statistically significant. Statistical analyses were performed using SPSS (version 27.0; IBM Corp) and Python (version 3.7).

This study was approved by the Institutional Review Board of the Chinese Academy of Medical Sciences and Peking Union Medical College (CAMS and PUMC-IEC-2022-022). Informed consent was not required due to the retrospective nature of the dataset, and all personal information and images were completely anonymized. We commit that all research data will be used for academic research purposes.

For the classification modeling, 8335 colposcopy images from 4 hospitals were selected for training and evaluation. These images included consisting of 2788 images of TZ1, 2663 images of TZ2, and 2884 images of TZ3. In the external validation study, 1335 cases were selected from 2 hospitals for model inference to predict the TZ types (Figure 5). The demographic characteristics of each participant, including age, colposcopic findings, and the distribution of each TZ type, are provided in Table 1. TZ types were found to be significantly associated with the participants’ age group distribution (P<.001). The mean age of all participants at the time of colposcopy was 37.87 (SD 9.99) years, with the mean age of the TZ3 group being 45.04 (SD 11.69) years.

Cervix detection with a bounding box was used for feature engineering. Proper cervix extraction improved classification accuracy. A total of 8335 cervix images were investigated and resized to 224×224 pixels before being input into the classification model. Out of these, 1663 images were used to evaluate the model’s performance.

Around 80% of the images were randomly selected as the training set, with the optimal weight parameter model selected during training being used to classify the images in the test set. The validation accuracy of the model gradually improved as the number of epochs increased, as shown in the validation plot (Figure 6). After approximately 100 epochs, both validation accuracy and training loss stabilized, reaching a peak validation accuracy of 83% at the 200th epoch. The validation accuracy fluctuated significantly during the first 25 epochs, it showed a general upward trend, while training loss rapidly decreased, indicating active learning and adjustment of the model. The model reached an optimal balance point and performed well after the 75th epoch when validation accuracy stabilized and training loss continued to decrease but with minimal changes. Overall, the model demonstrated rapid convergence in the early stages of training, with steady performance improvements in the later stages.

Our proposed model accurately classified cervical TZ types, as shown in Table 2. The highest classification accuracy of 83.97% and precision of 83.93% were achieved on the test set. For the 3 TZ types, the sensitivity was 84.74% for TZ1, 78.95% for TZ2, and 87.87% for TZ3, while the specificity was 89.99% for TZ1, 91.98% for TZ2, and 94.03% for TZ3. The detailed values for sensitivity, specificity, positive predictive value, and NPV are presented in Table 3. According to the classification performance of TZ types, the sensitivity and NPV for TZ2 were significantly lower.

The SOTA deep-learning networks for interpreting colposcopy images were selected to train on our dataset for comparison with our model. Table 2 presents the experimental results of the test set based on various evaluation metrics. The traditional ResNet50 and Vision Transformer models performed poorly in terms of accuracy and precision, achieving only 59.22% and 57.84%, and 59.46% and 57.64%, respectively. VGG16, with its relatively low-density network structure, performed better than ResNet50 in feature extraction. In addition, we selected 2 lightweight networks for comparison, EfficientNet and ShuffleNet, with higher accuracy of 62.06% and 74.47%. The proposed model achieved the highest accuracy of 83.97% among SOTA models.

To enhance AI’s clinical guidance, we used a pretrained FastSAM model to segment the TZ region by visualizing the new SCJ. A portion of the training data was used to annotate the new SCJ (Figure 7A), excluding TZ3, as the new SCJ is located in the endocervix. The segmentation model predicted the negative foreground region indicating the area outside the target TZ region (Figure 7B). The region boundary of the mask prediction was the predicted new SCJ (Figure 7C). In the test set, the overall recall and mAP50 (mean average precision at 50% Intersection over Union) of predicted mask were 0.78 and 0.75, respectively.

An independent dataset of 1335 cases was used to evaluate the generalization of the TZ classification model, distributed as 594 cases of TZ1, 604 cases of TZ2, and 137 cases of TZ3. The overall classification accuracy was 79.33%. The model’s predicted sensitivity for TZ1, TZ2, and TZ3 was 77.3% (95% CI 73.9%-80.6%), 81.1 (95% CI 78.0%-82.3%), and 80.3% (95% CI 73.7%-86.9%), respectively, while the specificity was 94.2% (95% CI 92.4%-95.8%), 83.3% (95% CI 80.6%-85.9%), and 90.7% (95% CI 89.1%-92.3%). The classification performance is presented in Figure 8 and Table 4.

In this study, we proposed a classification and identification model that facilitates clinical colposcopy examinations of cervical TZ. Our AI model achieved an accuracy of 79.33% in the task of classifying the three types of TZ in the external dataset. The sensitivity of TZ3 was 80.3%, and the specificity was 90.7%, which were satisfactory. TZ3 could be accurately differentiated, which further decreased the chance of missed diagnoses of high-grade lesions [21], and it was valuable for guiding ECC and recommending appropriate treatments. The AI model had increased sensitivity on TZ2 and lower sensitivity on TZ1 compared to the performance on the test set, which might be attributed to the similarity between TZ1 and TZ2. According to the standard terminology of the IFCPC, TZ1 and TZ2 were both visible either with or without the assistance of an endocervical speculum [22]. TZ2 was nearly entirely visible in some cases, while the SCJ was only visible at the endocervical canal’s border. Therefore, it was difficult to differentiate it from TZ1, which frequently required expert evaluation in strict accordance with established guidelines. In clinical practice, the treatment management for TZ1 and TZ2 is generally similar [6]. Our AI model demonstrated noninferior performance in classifying them and could be used to assist less experienced or inexperienced colposcopists in the colposcopic examination process. The overall model accuracy from the validation study was similar to the performance during model training, indicating that it is capable of generalizing fairly well.

In terms of AI techniques, we proposed a method based on the variant MobileNetV3 architecture for cervical TZ classification and FastSAM for segmenting TZ in colposcopy. Currently, only a few deep learning–based models have been developed to classify cervical TZ from colposcopy images. Dash et al [23] conducted a TZ segmentation and classification model based on colposcopy images from the IARC image bank. Similarly, Cao et al [24] developed a high-performance, deep learning model based on retrospective image data collected from one hospital [24]. Comparatively, our method was based on a multicenter study with a more diverse dataset of colposcopy images. While this diversity added to the challenge of feature extraction, the classification model showed better accuracy. Second, the proposed segmentation model could precisely annotate the new SCJ and indicate the approximate location of TZ, which could greatly assist junior colposcopists in selecting biopsy sites. Effective biopsy prioritization focused on the most severe lesions, particularly those within the TZ. Compared with the previous study [23], our method delineated a new SCJ rather than the original SCJ, which can provide more effective assistance and insights for colposcopic clinical examination and treatment. Furthermore, conducted external validation for the TZ classification model for the first time, and our model demonstrated strong stability and generalization performance.

In resource-limited settings, colposcopists demonstrated significantly lower clinical skills and performance than those observed in our study. The results of a study evaluating the colposcopic abilities of colposcopists in underserved Chinese communities indicated that colposcopists had a mere 22% accuracy specifically for TZ3 [8]. Similarly, in another study, colposcopists’ clinical diagnostic abilities were assessed before and after intensive training [9]. It was found that junior colposcopists had only 49.1% accuracy in classifying TZs. Despite the notable improvement, their accuracy remained below optimal levels at 68.6% following training. In addition, a comparable study conducted in Europe demonstrated that junior colposcopists were only able to detect TZs with a 55% accuracy rate [25]. According to these findings, colposcopists at less-experienced levels generally exhibited inadequate colposcopic performance. However, AI might offer a promising solution to enhance colposcopic capabilities and clinical decision-making confidence. Compared to these studies, our TZ classification model demonstrated strong performance with a high accuracy of 83.97% in the test set and achieved much higher sensitivity in predicting three TZ types in the validation study. Based on the results of previous studies, our model was more effective at stand-alone classification than that of junior colposcopists. Therefore, the method presented in this study accurately identifies TZ in colposcopy images, providing a valuable reference for colposcopists when making clinical decisions. The findings from this study supported the potential of the proposed AI-based TZ identification method as a promising adjunct tool for colposcopic examinations, particularly when integrated with AI colposcopic diagnostic systems and digital colposcopes. Dynamic digital imaging with AI assistance enhances the objectivity of colposcopic examinations and might address the diagnostic subjectivity of less experienced colposcopists. When AI-guided digital colposcopists are deployed in resource-limited health care settings, colposcopists will be able to receive intuitive and accessible guidance on clinical features from AI during the colposcopic examination in real time, supporting less-experienced colposcopists in improving their overall colposcopy skills. Combining AI results with colposcopist assessments helps reduce diagnostic bias, improve colposcopic examination capability, and narrow the gap with resource-rich areas. It addresses the minimization of missed diagnoses, the facilitation of early detection of lesions, the reduction of the risk of CIN progression, and the reduction of the burden of cervical cancer [26]. The application of AI will be integral to improving the quality of colposcopy services in low-resource settings.

The development and application research of AI-guided colposcopy models has emphasized their auxiliary value in clinical practice [14,27]. With the advent of innovations in AI algorithms, these models rely predominantly on mainstream CNNs coupled with transfer learning, as a method of unsupervised learning for image feature learning [28,29]. In some AI models, object detection is used as a metric for identifying lesion areas; however, it remains limited to binary classification tasks. Nevertheless, the multiclassification of normal cervix, low-grade squamous intraepithelial lesion, and high-grade squamous intraepithelial lesion plus poses significant challenges to these models. A clinical perspective suggested that acetowhite changes within TZ have a higher likelihood of association with CIN lesions than those outside of TZ. While the lesion extends across the new SCJ, a biopsy should be taken within the TZ or at the new SCJ. Therefore, by incorporating the TZ-type information as a weight in the model, the AI could be able to perform a more effective feature engineering process within the TZ. It is anticipated that the AI model will markedly enhance the capability of lesion detection, thus eliminating the current limitation in which precise identification is confined to CIN2+ lesions. Our proposed AI system delineated the new SCJ on high-resolution colposcopy images, and it enabled the approximate location of the TZ to be visualized more intuitively, thereby guiding more effective colposcopy examination and biopsy procedures.

For applying AI techniques in limited-resource settings, we proposed a model based on the MobileNetV3 architecture for cervical TZ classification in colposcopy images. Since Google introduced MobileNet in 2017 [30] as a lightweight CNN architecture, it has gained significant attention for its efficiency and accuracy. A lightweight AI model is more energy efficient and requires significantly fewer computational resources than a large-scale AI model, aligning with sustainable AI practices. It has been designed to operate efficiently on battery-operated devices, making it especially suitable for deployment in remote regions with limited power availability and internet access. By minimizing the reliance on high-cost hardware and extensive cloud infrastructure, lightweight AI models enable resource optimization in low-resource settings and facilitate equitable access to AI. Our AI model achieved efficient computation and robust classification performance on portable devices and can be applied actively in a variety of clinical settings to validate its generalization ability. A further economic evaluation is required to support the decision to adopt novel technologies in screening strategies.

There are several limitations to this study. First, the scale of the dataset we included was limited, which restricted the ability to support training deeper or more complex deep neural networks. However, the quality and standardization of image acquisition were assured, and the appropriate network depth was chosen to maximize feature extraction. It is necessary to obtain more high-quality colposcopy images with endocervix expansion by auxiliary instruments per case to train more complex networks that can extract additional features and potentially improve classification accuracy in TZ2 and TZ3. To address the suboptimal performance in TZ2 classification, expanding the dataset for TZ2 in future studies could enable the model to capture more distinct texture features that differentiate it from TZ1 and TZ3. The model could be further enhanced by applying edge detection techniques in regions where it is difficult to distinguish TZ2 from TZ3 and to highlight subtle morphological changes at the new SCJ. Second, the external validation was limited to evaluating the AI results alone. Further research is needed to assess the impact of AI-aided colposcopists in the classification of TZ. A prospective clinical trial is further needed to investigate whether the method can be applied to real-world colposcopy clinics. In addition, although we collected the validation datasets from two different hospitals, the scale of the external validation dataset is relatively limited, particularly when it comes to representing clinical scenarios that involve different colposcope vendors and lighting conditions. Future studies will focus on expanding both the size and scope of the dataset, incorporating data from more regions for more comprehensive validation. Furthermore, transfer learning or domain adaptation techniques may be used to improve the model’s robustness and adaptability to various imaging conditions. Finally, the TZ classification lacks an absolute subjective gold standard but is guided by colposcopic examination criteria and expert consensus. However, the colposcopy expert panel from tertiary hospitals conducted the “ground truth” categories for this study in accordance with established guidelines from IFCPC. In light of the subjectivity inherent in TZ classification, using only single-image modality data for model development is limited. Model optimization in the future may implement multimodal learning, which incorporates image data with associated clinical factors (eg, age, menstrual status, gravidity, and parity) to reduce the subjectivity of the model and enhance TZ feature discrimination. Furthermore, the construction of a knowledge graph based on existing consensus and colposcopic examination guidelines will help the model adhere to various rules or logic during the learning process.

In this study, an accurate AI-based method was developed and evaluated for automatically identifying cervical TZ using colposcopy images. The proposed method was the first application of a lightweight CNN for cervical feature extraction and applied a general segmentation model for TZ delineation among multicenter images, achieving commendable classification accuracy on TZ. The proposed approach has the potential to adapt to various colposcopy clinical environments and improve AI-guided colposcopy practice.