Cervical Pre-cancer Classification Using MLP Based on Hybrid Features from GLCM, LBP, and MobileNetV2

Abstract

The early and accurate diagnosis of cervical intraepithelial neoplasia lesions (CIN), particularly in a resource-limited environment, is paramount in helping to control the rising epidemic of cervical cancer. This research offers a hybrid classification model that merge texture features like Gray Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP), alongside semantic features from MobileNetV2. These features, after being extracted, are merged and supplied to a Multilayer Perceptron (MLP) for multiclass classification into Normal, CIN1, CIN2, or CIN3. The model was trained and evaluated using a 5-fold stratified cross-validation technique on an IARC dataset that contains 200 cases of colposcopy images. The experimental results illustrate that the model developed with a stratified k-fold cross-validation performed consistently well with high performance, average accuracy reported as 86.75% ± 2.62% and Cohen's kappa 0.7963 ± 0.0524 showed substantial to almost perfect in agreement across folds. The best performance was recorded for Fold 4 achieving 90.31% accuracy, while maintaining robust F1-scores across all classes.  This hybrid approach offers a promising direction for developing efficient and accurate computer-aided diagnosis (CAD) systems for cervical lesion classification.