Classification of Cervical Intraepithelial Neoplasia Based on Combination of GLCM and L*a*b* on Colposcopy Image Using Machine Learning

Colposcopy can potentially improve diagnostic ca-pabilities to guide cervical biopsies, contributing to improved cervical cancer screening. The previous study compared diagnostic accuracy in histologic CIN between senior and junior col-poscopists, the senior colposcopists showing higher sensitivity but lower specificity than the junior. This research has significantly contributed to categorizing CIN1, CIN2, and CIN3 by utilizing features extracted from GLCM and L*a*b* color spaces, using machine learning techniques to aid cervical pre-cancer diagnosis. Three machine learning tools were random forest, decision tree, and extra trees. The results of this study have shown that the extra tress classifier has the best performance value by using six GLCM features: contrast, correlation, energy, homogeneity, dissimilarity, asm, and one L*a*b* feature, namely median_1. The performance results have improved to 0.98 accuracy, 0.97 sensitivity, and 0.98 specificity. Processing CIN1 data on GLCM features took an average of 20.70 s and L*a*b* for 21.74 s. The processing of CIN2 data on GLCM features is 27.57 s, and L*a*b* is 27.37 s. The processing of CIN3 data on GLCM is 31.22 s, and L*a*b* is 31.38 s. In aggregate, the lowest average processing time using the GLCM features was 26.20 s. Based on the testing results, this research produces an FNR of 0.015, which indicates that this research in the future, when applied to medical care, has a slight possibility of diagnostic errors. Abnormal vascular features such as punctation and mosaic are the only significant if visible in limited acetowhite areas, making it challenging for researchers to select more appropriate features.