Comparison of performance of cervical cancer grading based on acetowhite areas

Abstract

Cervical cancer ranks fourth globally in terms of both incidence and mortality among women, making timely diagnosis essential for effective treatment. Although the acetowhite regions and their margins are important for cervical cancer staging, their potential for automated cancer grading remains underexplored. This study aimed to enhance diagnostic accuracy and grading precision by effectively analyzing the acetowhite region and its surroundings. Using four classifiers (Logistic Regression(LR), Random Forest(RF), XGBoost(XGB), and Support Vector Machine(SVM)), 464 cervical images (228 atypical and 236 positive cases) were analyzed. From a set of 75 features, the classifiers identified the top 5 based on feature importance. Receiver Operating Characteristic (ROC) analysis yielded the following precisions for models trained with masks containing only the acetowhite lesion: LR 0.80 (CI 95% 0.70-0.90), SVM 0.83 (CI 95% 0.75-0.92), RF 0.79 (CI 95% 0.69-0.89), XGB 0.66 (CI 95% 0.55-0.77). For models trained with masks including the acetowhite lesion and a 10-pixel margin: LR 0.79 (CI 95% 0.70-0.88), SVM 0.87 (CI 95% 0.78-0.95), RF 0.86 (CI 95% 0.77-0.94), XGB 0.84 (CI 95% 0.75-0.93). Our findings indicate that including a 10-pixel margin around acetowhite lesions improves classifier performance, suggesting its advantage in the automated classification of cervical images.

Keywords: Acetowhite; Cervical cancer; Classification; Machine learning.

Fig. 1

Examples of colposcopic images: An example of colposcopic images. All images were taken after acetic acid solution application; (a) Colposcopic image labeled Atypical; (b) Colposcopic image labeled Positive 1; (c) Colposcopic image labeled Positive 2.

Fig. 2

Flow chart of the research: (Exp 1) From the input images, the model gets the mask images that only represent areas of acetowhite lesion. The mask images are used to extract radiomics features. Going through feature extraction and feature selection, five features were selected for cervical cancer classification. Four different kinds of classifiers were trained using the selected features, then classified the test images and produced results. (Exp 2) works just as Exp 1, except that mask images used at Exp 2 were expanded 10 pixels to the margin of the lesion.

Fig. 3

Importance of 5 features selected via RFE for acetowhite areas.

Fig. 4

Importance of 5 features selected via RFE for areas with an additional 10 pixels from the acetowhite region.