Evaluasi Performa Efficientnetv2 Dalam Klasifikasi Citra Penyakit Lesi Kulit

Nurdiansyah, Hilmi Rahman (2025) Evaluasi Performa Efficientnetv2 Dalam Klasifikasi Citra Penyakit Lesi Kulit. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5002211075-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Penyakit kulit menular seperti monkeypox, chickenpox, dan measles memiliki kemiripan visual pada citra lesi, sehingga menyulitkan proses diagnosis dini secara akurat. Penelitian ini menggunakan arsitektur EfficientNetV2-B0 untuk mengklasifikasikan citra lesi kulit ke dalam empat kelas, termasuk kulit normal. Dua strategi pelatihan diterapkan, yaitu fine-tuning dengan bobot pralatih dari ImageNet dan pelatihan dari awal (from scratch) tanpa bobot awal. Selain itu, penelitian ini juga menganalisis pengaruh variasi hyperparameter berupa jumlah dense unit (64, 128, 256) dan jenis optimizer (Adam, AdamW, SGD) terhadap performa model. Dataset terdiri dari 477 citra berlabel yang telah diseimbangkan dan dibagi ke dalam rasio 80:10:10 untuk pelatihan, validasi, dan pengujian. Augmentasi citra dilakukan secara on-the-fly untuk meningkatkan variasi data. Sebanyak 18 skenario eksperimen dilakukan berdasarkan kombinasi strategi pelatihan dan konfigurasi arsitektur. Hasil menunjukkan bahwa model dengan pendekatan fine-tuning cenderung menghasilkan metrik evaluasi yang lebih tinggi dibandingkan pelatihan from scratch, terutama pada data uji. Namun demikian, seluruh konfigurasi masih menunjukkan gejala overfitting dan fluktuasi performa antar-epoch, terutama pada data validasi. Temuan ini mengindikasikan bahwa baik pendekatan fine-tuning maupun pelatihan from scratch belum berhasil menghasilkan model yang dapat diandalkan, karena seluruh konfigurasi masih menunjukkan gejala overfitting dan fluktuasi performa. Hal ini menunjukkan bahwa strategi pelatihan dan kombinasi hyperparameter yang digunakan dalam eksperimen ini belum cukup efektif untuk mencapai kemampuan generalisasi yang baik.
========================================================================================================================
Infectious skin diseases such as monkeypox, chickenpox, and measles exhibit high visual similarity in lesion images, making early and accurate diagnosis challenging. This study employs the EfficientNetV2-B0 architecture to classify skin lesion images into four categories, including normal skin. Two training strategies were applied: fine-tuning using pretrained weights from ImageNet and training from scratch without pretrained weights. Additionally, the study investigates the impact of hyperparameter variations, including the number of dense units (64, 128, 256) and optimizer types (Adam, AdamW, SGD), on model performance. The dataset consists of 477 labeled and balanced images, split in an 80:10:10 ratio for training, validation, and testing. On-the-fly data augmentation was used to enhance data variability. A total of 18 experimental scenarios were conducted based on combinations of training strategies and model configurations. The results show that fine-tuning generally produced higher evaluation metrics on test data compared to training from scratch. However, all configurations still exhibited signs of overfitting and performance fluctuations across epochs, particularly on validation data. These findings indicate that neither training strategy successfully produced a reliable model. The training approach and hyperparameter settings used in this study were not sufficiently effective in achieving stable generalization performance.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	klasifikasi citra, EfficientNetV2, Monkeypox, penyakit kulit menular,deep learning. Image classification, Monkeypox, EfficientNetV2, Infectious skin disease, Deep learning.
Subjects:	Q Science > QA Mathematics > QA76.6 Computer programming. Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science)
Divisions:	Faculty of Mathematics, Computation, and Data Science > Mathematics > 44201-(S1) Undergraduate Thesis
Depositing User:	Hilmi Rahman Nurdiansyah
Date Deposited:	04 Aug 2025 04:30
Last Modified:	04 Aug 2025 04:30
URI:	http://repository.its.ac.id/id/eprint/123726

Actions (login required)

View Item