Experimenting with the Hyperparameter of Six Models for Glaucoma Classification

Ilham, Muhamad and Prihantoro, Angga and Perdana, Iqbal Kurniawan and Magdalena, Rita and Saidah, Sofia (2023) Experimenting with the Hyperparameter of Six Models for Glaucoma Classification. Jurnal Ilmiah Teknik Elektro Komputer dan Informatika (JITEKI), 9 (3). pp. 571-584.

[thumbnail of 6-Experimenting with the Hyperparameter of Six Models for Glaucoma Classification.pdf] Text
6-Experimenting with the Hyperparameter of Six Models for Glaucoma Classification.pdf

Download (1MB)


Glaucoma is characterized by optic nerve damage, and can potentially lead to blindness, often presenting with no obvious symptoms in most affected individuals. As a result, a large proportion of those affected remain undiagnosed, making early detection crucial for effective treatment. Numerous studies have been conducted to develop glaucoma detection systems. In this particular study, a glaucoma detection system using the CNN method was developed. The contribution of this research is to conduct hyperparameter experiments on AlexNet, Custom Layer, MobileNetV2, EfficientNetV1, InceptionV3, and VGG19 models on the RIM-ONE DL dataset with a total of 933 images that have been augmented. Hyperparameter experiments were conducted to determine the most optimal parameters for each model, specifically testing batch size, learning rate, and optimizer. The batch sizes used were 64, 128, 256, and 512. The learning rates used are 0.1, 0.001, 0.0001, and 0.00001. The optimizers to be tested are Nadam, Adam, and RMSProp. The hyperparameter optimization process yielded the optimal parameters for each model. However, it is important to note that the MobileNetV2, InceptionV1, and VGG19 models exhibited signs of overfitting in the training graph results. Among the models, the custom layer model achieved the highest accuracy of 93%, while InceptionV3 attained the lowest accuracy at 83.5%. Model testing was conducted using data from the Cicendo Eye Hospital, and the RIM-ONE DL testing dataset which totals 200 images. Based on the testing results, it was found that InceptionV3 outperformed the other models in predicting images accurately. Therefore, the study concluded that high accuracy in training does not necessarily indicate superior performance in testing, particularly when limited variation exists in the training dataset.

Item Type: Artikel Umum
Subjects: T Technology > TK Electrical engineering. Electronics Nuclear engineering
Divisi / Prodi: Faculty of Industrial Technology (Fakultas Teknologi Industri) > S1-Electrical Engineering (S1-Teknik Elektro)
Depositing User: M.Eng. Alfian Ma'arif
Date Deposited: 28 Jul 2023 01:10
Last Modified: 28 Jul 2023 01:10
URI: http://eprints.uad.ac.id/id/eprint/43712

Actions (login required)

View Item View Item