Skripsi
KLASIFIKASI PRA-KANKER SERVIKS BERBASIS FITUR HIBRIDA DENGAN CITRA KOLPOSKOPI MENGGUNAKAN MULTI LAYER PERCEPTRON DAN DEEP NEURAL NETWORK
Cervical cancer remains one of the leading causes of cancer-related mortality among women worldwide, particularly in resource-limited settings where access to routine screening and expert medical evaluation is restricted. Early detection of Cervical Intraepithelial Neoplasia (CIN) through colposcopy image analysis is critical to prevent the progression of pre-cancerous lesions into invasive cervical cancer. This study proposes a hybrid feature-based classification approach for cervical pre-cancer using two model schemes. The first scheme combines conventional texture feature extraction methods, namely Gray Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP), with deep features extracted from MobileNetV2, subsequently classified using a Multilayer Perceptron (MLP). The second scheme integrates Gray Level Run Length Matrix (GLRLM) texture features, Color Moments, and deep features from EfficientNetB3 and MobileNetV3, classified using a Deep Neural Network (DNN). The dataset consists of 200 colposcopy images obtained from the International Agency for Research on Cancer (IARC), categorized into four classes: Normal, CIN1, CIN2, and CIN3. Image augmentation was applied to address data scarcity, and model evaluation was conducted using stratified 5-fold cross-validation with accuracy, precision, recall, F1-score, and Cohen's Kappa as performance metrics. Experimental results demonstrate that the MLP-based hybrid feature model achieved a mean accuracy of 86.75% ± 2.62% and a mean Cohen's Kappa of 0.7963, while the DNN-based hybrid feature model achieved a mean accuracy of 82.67% ± 3.65% and a mean Cohen's Kappa of 0.7466. Both models consistently exhibited strong performance on the Normal and CIN3 classes, whereas CIN2 classification remained the most challenging due to overlapping visual and textural characteristics with adjacent lesion grades. This study demonstrates that the hybrid feature fusion approach holds significant potential as a computer-aided decision support system for objective and automated early screening of cervical pre-cancerous lesions. Keywords: Cervical pre-cancer, CIN, colposcopy image, hybrid features, MLP, DNN, GLCM, LBP, GLRLM, Color Moments, MobileNetV2, MobileNetV3, EfficientNetB3
No other version available