Klasifikasi Diabetic Retinopathy melalui Analisis Citra Fundus Menggunakan Deep Learning dengan Visualisasi Heatmap

Larasati, Sharfina Nabila (2025) Klasifikasi Diabetic Retinopathy melalui Analisis Citra Fundus Menggunakan Deep Learning dengan Visualisasi Heatmap. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5023211055-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Diabetes mellitus merupakan penyebab utama komplikasi mikrovaskular, salah satunya adalah retinopati diabetik (DR), yang dapat menyebabkan kebutaan permanen jika tidak ditangani. DR diklasifikasikan menjadi Non-Proliferative DR (NPDR) dengan tingkat keparahan normal, mild, moderate, dan severe, serta Proliferative DR (PDR). Penilaian tingkat keparahan ini bergantung pada keberadaan dan jumlah lesi seperti microaneurysms, hemorrhages, dan exudates yang terdeteksi pada citra fundus retina. Proses diagnosis secara manual oleh dokter memerlukan waktu dan keahlian tinggi. Oleh karena itu, penelitian ini bertujuan untuk meningkatkan efisiensi dan akurasi diagnosis dengan menggunakan sistem berbasis teknologi komputer yang efisien dan akurat. Model yang digunakan adalah EfficientNetB3 dengan pendekatan transfer learning dua tahap, yaitu pre-training dan fine-tuning. Dataset yang digunakan merupakan gabungan dari APTOS 2019 dan IDRiD. Serta data primer dari RS Mata Undaan yang digunakan untuk pengujian model. Ketiga dataset dilakukan pre-processing berupa shape normalization, resize, color normalization, CLAHE, serta augmentasi data untuk menyeimbangkan distribusi kelas. Visualisasi Grad-CAM digunakan untuk membantu interpretasi hasil klasifikasi dengan menyoroti area penting pada citra fundus. Hasil evaluasi menunjukkan bahwa model EfficientNetB3 yang telah dilakukan fine tuning mampu mengklasifikasikan tingkat keparahan DR secara akurat, dengan akurasi sebesar 88,32%, recall 88,02%, dan spesifisitas 97,07%. Visualisasi Grad-CAM juga menunjukkan bahwa model mampu fokus pada area yang relevan secara klinis. Dengan hasil ini, sistem yang dikembangkan diharapkan dapat menjadi alat bantu diagnosis yang efisien, akurat, dan mendukung pengambilan keputusan medis dalam mendeteksi dan mengklasifikasikan DR secara tepat dan cepat.
====================================================================================================================================
Diabetes mellitus is a major cause of microvascular complications, often appearing as diabetic retinopathy (DR), which is one of the leading causes of preventable blindness. DR is clinically divided into Non-Proliferative DR (NPDR) which includes mild, moderate, and severe stages and Proliferative DR (PDR). The severity level is determined based on the presence and appearance of specific retinal lesions such as microaneurysms, hemorrhages, and exudates, which can be identified in fundus images. However, manual diagnosis by ophthalmologists can be time-consuming and requires significant clinical expertise. To address this challenge, this study proposes a deep learning-based system to automatically classify the severity levels of DR from fundus images. The system is built using the EfficientNetB3 architecture and applies a two-stage transfer learning approach: pre- training and fine-tuning. The dataset used includes a combination of APTOS 2019 and IDRiD fundus images, as well as primary data from Undaan Eye Hospital, which was used for model testing. These images are subjected to a detailed preprocessing pipeline, which includes shape normalization, resizing, color normalization, contrast enhancement using Contrast-Limited Adaptive Histogram Equalization (CLAHE), and data augmentation to address class imbalance. To improve the interpretability of the model, Grad-CAM visualization is used to produce heatmaps that highlight important regions in the fundus images that influence the classification. Evaluation results show that the fine-tuned EfficientNetB3 model achieves strong classification performance, with an accuracy of 88.32%, recall of 88.02%, and specificity of 97.07%. The Grad-CAM heatmaps also confirm that the model focuses on clinically meaningful features. This reliable and explainable system is expected to assist clinicians in making faster, more consistent, and accurate decisions when diagnosing the severity of diabetic retinopathy.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Diabetes Mellitus, Retinopati Diabetik (DR), Deep Learning, EfficientNetB3, Grad-CAM, Diabetic Retinopathy (DR)
Subjects:	R Medicine > R Medicine (General) > R858 Deep Learning R Medicine > RE Ophthalmology R Medicine > RE Ophthalmology > RE48 Eye--Diseases. Ophthalmoscopy. T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Electrical Technology > Biomedical Engineering > 11410-(S1) Undergraduate Thesis
Depositing User:	Sharfina Nabila Larasati
Date Deposited:	05 Aug 2025 01:32
Last Modified:	05 Aug 2025 01:32
URI:	http://repository.its.ac.id/id/eprint/127344

Actions (login required)

View Item