KLASIFIKASI TINGKAT KEPARAHAN RETINOPATI DIABETIK MENGGUNAKAN STACKING ENSEMBLE DENGAN EFFICIENTNET V2-S, RESNET50, DAN DENSENET121

Basori, Muhammad Iqmal (2026) KLASIFIKASI TINGKAT KEPARAHAN RETINOPATI DIABETIK MENGGUNAKAN STACKING ENSEMBLE DENGAN EFFICIENTNET V2-S, RESNET50, DAN DENSENET121. Undergraduate thesis, UPN Veteran Jawa Timur.

	Text (Cover) 21081010297.-cover.pdf Download (819kB)
	Text (Bab 1) 21081010297.-bab1.pdf Download (253kB)
	Text (Bab 2) 21081010297.-bab2.pdf Restricted to Repository staff only until 19 January 2029. Download (907kB)
	Text (Bab 3) 21081010297.-bab3.pdf Restricted to Repository staff only until 19 January 2029. Download (2MB)
	Text (Bab 4) 21081010297.-bab4.pdf Restricted to Repository staff only until 19 January 2029. Download (10MB)
	Text (Bab 5) 21081010297.-bab5.pdf Download (243kB)
	Text (Daftar Pustaka) 21081010297.-daftarpustaka.pdf Download (186kB)
	Text (Lampiran) 21081010297.-lampiran.pdf Restricted to Repository staff only until 19 January 2029. Download (351kB)

Abstract

Diabetic retinopathy is a microvascular complication of diabetes mellitus that can cause permanent blindness if Not detected and treated early. Manual detection requires considerable time and is limited by the availability of expert personnel, making Computer-Aided Diagnosis (CAD) systems based on deep learning an important solution for mass screening. This research aims to develop a Diabetic retinopathy severity classification system using Stacking Ensemble that integrates three pre-trained CNN architectures: EfficientNetV2-S, ResNet50, and DenseNet121 as base learners with Gradient Boosting Classifier as the meta-learner. The dataset is a combination of APTOS 2019, IDRiD, and Messidor-2 with a total of 5,922 retinal fundus images that underwent preprocessing Stages of ROI Extraction, Gamma correction, CLAHE, and Resizing. Each base learner was trained on three dataset split scenarios (80:10:10, 70:15:15, and 60:20:20) with three sampling methods (Random sampling, Undersampling, and No Augmentation), resulting in a total of 36 models. Evaluation results show that EfficientNetV2-S achieved the best performance in the 70:15:15 scenario with Random sampling, achieving an Accuracy of 83.11%, Precision of 0.8304, recall of 0.8311, F1-Score of 0.8289, and QWK of 0.9298. The Stacking Ensemble method with the same configuration achieved an Accuracy of 82.78% and QWK of 0.9293. This research demonstrates that the Ensemble learning approach with Gradient Boosting meta-learner is capable of overcoming model performance degradation when the amount of training Data is reduced. This is achieved by leveraging complementary patterns from the three model architectures used.

Item Type:

Thesis (Undergraduate)

Contributors:

Contribution	Contributors	NIDN/NIDK	Email
Thesis advisor	Diyasa, I Gede Susrama Mas	NIDN0019067008	igsusrama.if@upnjatim.ac.id
Thesis advisor	Maulana, Hendra	NIDN1423128301	hendra.maulana.if@upnjatim.ac.id

Subjects:

Q Science > QA Mathematics > QA76.6 Computer Programming

Divisions:

Faculty of Computer Science > Departemen of Informatics

Depositing User:

Muhammad Iqmal Basori

Date Deposited:

19 Jan 2026 08:37

Last Modified:

18 Feb 2026 07:31

URI:

https://repository.upnjatim.ac.id/id/eprint/48761