JMIR Publications

ABSTRACT

Background:

Cervical cancer remains the fourth most prevalent cancer among women worldwide, and a leading cause of morbidity and mortality in low- and middle-income countries (LMICs). Early detection and treatment of precancerous lesions are critical in cervical cancer prevention. There is an urgent need for appropriate screening strategies in these areas, with novel screening technologies being the most preferable option. Colposcopy is a primary diagnostic tool used to identify cervical lesions and guide biopsies. The transformation zone (TZ) is the area where a stratified squamous epithelium develops from the metaplasia of simple columnar epithelium and is the most common site of precancerous lesions. When performing a colposcopy examination, it is crucial to identify the location and type of the transformation zone in order to improve the detection of cervical lesions.

Objective:

This study aims to present an artificial intelligent method for identifying the cervical TZ enhancing colposcopy examination and evaluate its potential clinical application.

Methods:

The study was retrospectively collected from 3,616 women who underwent colposcopy at six tertiary hospitals in China between 2019 and 2021. Anonymized digital records included baseline characteristics, colposcopic findings, and pathological outcomes. The proposed model is a lightweight neural network with multiscale feature enhancement capabilities, designed to classify the three types of TZ. It was employed by a series of multi-scale convolution layers based on MobileNetV3 framework, culminating in Squeeze-and-Excitation (SE) blocks for robust performance and regularization. The region of interest (ROI) in each high-resolution colposcopic image was automatically detected and segmented to ensure precise model training. TZ was segmented by FastSAM as a pioneering research for identifying the location of new squamocolumnar junction (SCJ). The data were randomly assigned (8:2) to a training set for developing the model and a test set for assessing the performance of TZ classifications. After the evaluation of model, an independent dataset was collected for model validation.

Results:

The optimal TZ classification model performed 83.97% for classification accuracy on test set, which achieved average precision of type1, type2, and type3 were 91.84%, 89.06%, and 95.62% respectively. The recall and mAP of TZ segmentation model were 0.69 and 0.68. The proposed model demonstrated outstanding performance in classifying TZ3, achieving a sensitivity of 87.87% and a specificity of 94.03%. Remarkably, the overall classification accuracy for TZ types achieved 79.33%, based on a comprehensive external dataset of 1,335 cases from two out of six hospitals.

Conclusions:

Our proposed AI-based identification system accurately classified the type of cervical TZs and delineate the location on multicenter colposcopic high-resolution images. It has the potential to enhance AI-guided colposcopy examination and accurately guided biopsy. It also could be a valued tool for assisting colposcopic diagnosis in clinical practice.