Optimasi Augmentasi Stylegan2 Dan Fine Tuning EfficientNetV2 Untuk Klasifikasi Tumor Otak Berbasi MRI

Pratama, Rifqi Zumadila (2026) Optimasi Augmentasi Stylegan2 Dan Fine Tuning EfficientNetV2 Untuk Klasifikasi Tumor Otak Berbasi MRI. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6025221007-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

Penelitian ini bertujuan untuk mengintegrasikan teknik augmentasi citra sintetis berbasis StyleGAN2 dan fine-tuning arsitektur EfficientNetV2 guna meningkatkan kinerja klasifikasi multikelas tumor otak berbasis MRI. Keterbatasan jumlah data berlabel serta variasi morfologi lesi sering menyebabkan model mengalami overfitting. Untuk mengatasinya, StyleGAN2 digunakan untuk mensintesis citra dengan fidelitas anatomi tinggi yang menyerupai pola radiologis nyata, sementara EfficientNetV2 dipilih karena rasio efisiensi komputasi dan akurasinya yang superior pada dataset medis. Penelitian ini juga menerapkan Explainable Artificial Intelligence (XAI) menggunakan Grad-CAM untuk memvisualisasikan atensi model, memastikan hasil prediksi dapat dipertanggungjawabkan secara klinis. Evaluasi dilakukan secara komprehensif dengan membandingkan skenario dengan data asli dibandingkan dengan skenario injeksi data sintetis dari berbagai metode GAN (DCGAN, FastGAN, dan StyleGAN2) pada berbagai arsitektur pengekstraksi fitur (EfficientNetV2, VGG-16, dan ResNet). Hasil pengujian menunjukkan bahwa kombinasi usulan utama, yakni EfficientNetV2 dan StyleGAN2, mencapai performa puncak dengan akurasi 99,00%, melampaui performa model yang menggunakan data asli sebesar 96,80%. Eksperimen komparatif juga membuktikan bahwa penambahan variasi data sintetis yang sama justru mendegradasi akurasi arsitektur klasik seperti VGG-16, menegaskan perlunya arsitektur modern untuk memproses data dari Generative AI. Visualisasi Grad-CAM secara konsisten memvalidasi bahwa model memusatkan atensinya secara presisi pada patologi lesi tumor, menjadikan kerangka kerja ini sangat potensial sebagai sistem pendukung keputusan diagnostik yang andal.
===================================================================================================================================
This study aims to integrate StyleGAN2-based synthetic image augmentation and fine-tuning of the EfficientNetV2 architecture to improve the performance of multiclass brain tumor classification using MRI images. The limited availability of labeled data and the morphological variability of lesions often cause models to suffer from overfitting. To address these challenges, StyleGAN2 is employed to synthesize images with high anatomical fidelity that closely resemble real radiological patterns, while EfficientNetV2 is selected for its superior balance between computational efficiency and classification accuracy on medical imaging datasets. In addition, this study incorporates Explainable Artificial Intelligence (XAI) through Gradient-weighted Class Activation Mapping (Grad-CAM) to visualize the model’s attention mechanisms, ensuring that prediction outcomes can be clinically interpreted and justified. A comprehensive evaluation is conducted by comparing scenarios using only original data against scenarios augmented with synthetic images generated by different GAN architectures, including DCGAN, FastGAN, and StyleGAN2, across multiple feature extraction networks such as EfficientNetV2, VGG-16, and ResNet. The experimental results demonstrate that the proposed combination of EfficientNetV2 and StyleGAN2 achieves the highest performance, reaching an accuracy of 99.00%, outperforming the baseline model trained solely on original data, which achieved 96.80% accuracy. Comparative experiments further reveal that injecting the same synthetic data variations can actually degrade the performance of classical architectures such as VGG-16, highlighting the importance of modern deep learning architectures in effectively leveraging Generative AI-produced data. Grad-CAM visualizations consistently validate that the model precisely focuses its attention on tumor lesion regions, indicating that the proposed framework has strong potential as a reliable clinical decision-support system for brain tumor diagnosis.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Augmentasi Data, EfficientNetV2, Grad-CAM, Klasifikasi Citra Medis, MRI, StyleGAN2, Tumor Otak, XAI, Brain Tumor, Data Augmentation, EfficientNetV2, Grad-CAM, Medical Image Classification, MRI, StyleGAN2, XAI.
Subjects:	R Medicine > R Medicine (General) > R858 Deep Learning R Medicine > RB Pathology
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 59101-(S2) Master Thesis
Depositing User:	Rifqi Zumadila Pratama
Date Deposited:	04 Aug 2026 06:09
Last Modified:	04 Aug 2026 06:09
URI:	http://repository.its.ac.id/id/eprint/143122

Actions (login required)

View Item