HAKIM, Nathasya Utami and Sugiharto, Aris and Wibawa, Helmie Arif (2026) PENINGKATAN PERFORMA MODEL EFFICIENTNETB5 MELALUI OPTIMASI HYPERPARAMETER UNTUK KLASIFIKASI TUMOR OTAK. Masters thesis, UNIVERSITAS DIPONEGORO.

	Text cover awal.pdf Download (124kB)
	Text COVER.pdf Restricted to Repository staff only Download (1MB)
	Text BAB I.pdf Download (297kB)
	Text BAB II.pdf Download (668kB)
	Text BAB III.pdf Restricted to Repository staff only Download (797kB)
	Text BAB IV.pdf Restricted to Repository staff only Download (628kB)
	Text BAB V.pdf Restricted to Repository staff only Download (292kB)
	Text DAFTAR PUSTAKA.pdf Download (260kB)
	Text LAMPIRAN.pdf Restricted to Repository staff only Download (1MB)

Abstract

Tumor otak merupakan penyakit serius yang membutuhkan diagnosis cepat dan akurat karena keterlambatan dapat menurunkan peluang keberhasilan terapi. Pendekatan konvensional sering menghadapi kendala dalam mengenali variasi bentuk, ukuran, dan intensitas tumor pada citra MRI sehingga berpotensi menghasilkan interpretasi yang berbeda antar ahli radiologi. Untuk meningkatkan akurasi identifikasi, penelitian ini mengoptimalkan arsitektur EfficientNetB5 melalui perbandingan empat metode hyperparameter tuning, yaitu Grid Search, Random Search, Bayesian Optimization, dan Particle Swarm Optimization (PSO). Proses penelitian meliputi preprocessing citra, normalisasi menggunakan preprocess input, augmentasi untuk mengurangi ketidakseimbangan kelas, pelatihan model, serta evaluasi melalui metrik akurasi, presisi, recall, dan f1-score. Dataset yang digunakan adalah Brain Tumor Classification milik Sartaj Bhuvaji dengan empat kelas: glioma, meningioma, pituitary tumor, dan no tumor. Hasil evaluasi menunjukkan bahwa seluruh metode menghasilkan performa tinggi dengan akurasi berkisar antara 95,74% hingga 96,55%. Bayesian Optimization memberikan kinerja terbaik dengan akurasi 96,55%, presisi 96,64%, recall 96,82%, dan f1-score 96,72%. PSO berada pada posisi berikutnya dengan akurasi 96,15%, diikuti Grid Search (95,94%) dan Random Search (95,74%). Selain itu, analisis hasil klasifikasi memperlihatkan bahwa seluruh metode mampu mengenali keempat kelas secara konsisten dengan tingkat kesalahan yang rendah. Temuan ini menegaskan bahwa pemilihan metode hyperparameter tuning berpengaruh signifikan terhadap kualitas dan stabilitas model, dengan Bayesian Optimization terbukti sebagai metode yang paling efektif untuk mengoptimalkan EfficientNetB5 pada tugas klasifikasi tumor otak berbasis citra MRI. Penelitian ini memberikan kontribusi empiris terhadap penerapan teknik optimasi modern dalam meningkatkan kinerja model deep learning pada domain citra medis.
Kata Kunci: Deep Learning, Klasifikasi Tumor Otak, Optimasi Hyperparameter, EfficientNetB5, Analisis Citra MRI

Brain tumors are critical medical conditions that require rapid and accurate diagnosis, as delays can significantly affect treatment outcomes and patient survival rates. Conventional diagnostic approaches often struggle to interpret variations in tumor shape, size, and intensity in MRI images, which may lead to inconsistencies among radiologists. To enhance diagnostic reliability, this study optimizes the EfficientNetB5 architecture by comparing four hyperparameter tuning methods: Grid Search, Random Search, Bayesian Optimization, and Particle Swarm Optimization (PSO). The research workflow includes image preprocessing, normalization using preprocess input, data augmentation to address class imbalance, model training, and performance evaluation using accuracy, precision, recall, and F1-score. The dataset used is the Brain Tumor Classification dataset by Sartaj Bhuvaji, consisting of four categories: glioma, meningioma, pituitary tumor, and no tumor. Experimental results show that all tuning methods achieve strong performance, with accuracy ranging from 95.74% to 96.55%. Bayesian Optimization produces the best results with 96.55% accuracy, 96.64% precision, 96.82% recall, and a 96.72% F1-score. PSO achieves the second-highest accuracy at 96.15%, followed by Grid Search (95.94%) and Random Search (95.74%). Additional analysis demonstrates that all methods consistently classify the four tumor types with low error rates. These findings highlight the significant impact of hyperparameter tuning strategies on model accuracy and stability, with Bayesian Optimization emerging as the most effective approach for optimizing EfficientNetB5 in MRI-based brain tumor classification. This study provides empirical evidence supporting the role of modern optimization techniques in enhancing deep learning performance for medical imaging tasks.
Keyword: Deep Learning, Brain Tumor Classification, Hyperparameter Optimization, EfficientNetB5, MRI Image Analysis

Actions (login required)

View Item