Severity-Based Classification of Diabetic Retinopathy Using MobileNetV3

Uttami, Ni Wayan Putri Satya (2026) Severity-Based Classification of Diabetic Retinopathy Using MobileNetV3. Undergraduate thesis, Universitas Pendidikan Ganesha.

[img] Text (COVER)
COVER-2215091065_signed.pdf
Download (1MB)
[img] Text (ABSTRAK)
ABSTRAK-2215091065_signed.pdf
Download (286kB)
[img] Text (BAB 1 PENDAHULUAN)
PENDAHULUAN-2215091065_signed.pdf
Download (403kB)
[img] Text (BAB 2 KAJIAN TEORI)
TEORI-2215091065_signed.pdf
Restricted to Repository staff only
Download (767kB) | Request a copy
[img] Text (BAB 3 METODELOGI PENELITIAN)
METHOD-2215091065_signed.pdf
Restricted to Repository staff only
Download (713kB) | Request a copy
[img] Text (BAB 4 HASIL DAN PEMBAHASAN)
HASIL-2215091065_signed.pdf
Restricted to Repository staff only
Download (1MB) | Request a copy
[img] Text (BAB 5 PENUTUP)
KESIMPULAN-2215091065_signed.pdf
Restricted to Repository staff only
Download (394kB) | Request a copy
[img] Text (DAFTAR PUSTAKA)
DAFPUS-2215091065_signed.pdf
Download (416kB)
[img] Text (LAMPIRAN)
LAMPIRAN-2215091065_signed.pdf
Download (1MB)

Abstract

Diabetic retinopathy is a retinal complication of diabetes mellitus that can lead to vision loss if not detected early. This study aims to design and evaluate a severity-based diabetic retinopathy classification scheme using the MobileNetV3 architecture. The main dataset used was the Indian Diabetic Retinopathy Image Dataset (IDRiD), supported by controlled external-data experiments using DiaRetDB and APTOS. Several experimental schemes were tested, including direct RGB classification, CLAHE-based class balancing, MobileNetV3-Small and MobileNetV3-Large comparison, offline MobileNet feature classification, dual-backbone RETFound–MobileNetV3 feature fusion, and external-data addition. The model performance was evaluated using accuracy, Quadratic Weighted Kappa, macro AUC, macro F1-score, and weighted F1-score. The best experimental model was the Gated Feature Fusion model with All-Grade APTOS external augmentation at ratio 2.0, achieving an accuracy of 0.6505 and QWK of 0.7790. However, the CLAHE-Balanced RGB MobileNetV3-Large model was selected for mobile prototype implementation because it supports a single end-to-end image-to-prediction pipeline. This model achieved an accuracy of 0.6699 and was successfully converted into TensorFlow Lite and integrated into a Flutter-based Android prototype. The prototype demonstrates the feasibility of local preliminary diabetic retinopathy severity classification, although further clinical validation and device testing are required.

Item Type: Thesis (Undergraduate)
Uncontrolled Keywords: diabetic retinopathy, MobileNetV3, severity classification, TensorFlow Lite, Flutter
Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
Q Science > QA Mathematics > QA76 Computer software
Divisions: Fakultas Teknik dan Kejuruan > Jurusan Teknik Informatika > Program Studi Sistem Informasi (S1)
Depositing User: Ni Wayan Putri Satya Uttami
Date Deposited: 21 Jul 2026 08:09
Last Modified: 21 Jul 2026 08:09
URI: http://repo.undiksha.ac.id/id/eprint/31133

Actions (login required)

View Item View Item