Skripsi
PEMANFAATAN KECERDASAN BUATAN GENERATIF UNTUK LESI PRA-KANKER SERVIKS DALAM MENINGKATKAN AKURASI PENGAMBILAN KEPUTUSAN KLINIS
Data scarcity and class imbalance remain significant challenges in developing deep learning models for medical image classification, including the classification of Cervical Intraepithelial Neoplasia (CIN). Limited data availability and imbalanced class distribution, particularly in minority classes such as CIN 2 and CIN 3, may affect the stability, generalization capability, and performance of multi-class classification models. Conventional augmentation techniques provide limited improvement in distributional diversity. This study investigates the use of StyleGAN2-ADA as a Generative Adversarial Network (GAN)-based augmentation method to generate synthetic images that approximate the distribution of real CIN data while helping improve class balance. The evaluation was conducted using two classification settings: models trained with real data only and models trained with a combination of real and augmented data. In addition, two experimental schemes were designed to analyze the effect of data quantity on generative quality, namely GAN training after data splitting and GAN training before data splitting. This study evaluated several deep learning classification models, including YOLOv11, Vision Transformer (ViT), EfficientNetV2-S, and ConvNeXt. The quality of synthetic images was assessed using the Fréchet Inception Distance (FID), while classification performance was evaluated using accuracy, precision, recall, and F1-score. The results indicate that lower FID values tend to be associated with more consistent improvements in classification performance, including better handling of class imbalance reflected by improved F1-score values. Among the evaluated models, EfficientNetV2-S achieved the best performance with an accuracy and F1-score of 88% after augmentation. These findings suggest that GAN-based augmentation can contribute to improving CIN classification performance while helping address data scarcity and class imbalance, provided that generative quality and experimental design are appropriately managed.
| Title | Edition | Language |
|---|---|---|
| KLASIFIKASI PENYAKIT GINJAL KRONIS DENGAN MENGGUNAKAN ALGORITMA MACHINE LEARNING | id | |
| PENGEMBANGAN DIALOG ADAPTIF MENGGUNAKAN KECERDASAN BUATAN GENERATIF PADA GIM VISUAL NOVEL | id |