Electronic Theses and Dissertation

Retinopati diabetik merupakan komplikasi serius akibat diabetes melitus yang menyebabkan kerusakan pembuluh darah retina dan berpotensi menimbulkan kebutaan. Deteksi dini melalui segmentasi citra fundus menjadi langkah penting untuk membantu diagnosis penyakit ini. Penelitian ini bertujuan mengembangkan model segmentasi citra fundus retina menggunakan arsitektur R2AU-Net dengan penerapan Contrast Limited Adaptive Histogram Equalization (CLAHE) sebagai pra-pemrosesan guna meningkatkan kontras citra. Proses CLAHE dilakukan dengan parameter clipLimit 6.0 dan tileGridSize 8×8, kemudian dataset dibagi menjadi dua kategori, yaitu tanpa pra-proses CLAHE dan dengan pra-proses CLAHE. Model dilatih menggunakan optimizer Stochastic Gradient Descent dengan learning rate 10⁻², epoch 100, dan batch size 4, serta dievaluasi menggunakan metrik Pixel Accuracy, Dice Coefficient, dan Intersection over Union (IoU). Hasil menunjukkan bahwa model dengan pra-proses CLAHE menghasilkan Pixel Accuracy 0,8833, Dice Coefficient 0,4519, dan IoU 0,3440, sedangkan tanpa CLAHE memperoleh Pixel Accuracy 0,9261, Dice Coefficient 0,3742, dan IoU 0,2746. Peningkatan nilai Dice dan IoU menunjukkan bahwa penerapan CLAHE berhasil memperbaiki kualitas segmentasi dengan menonjolkan struktur pembuluh darah halus, sehingga kombinasi CLAHE dan R2AU-Net efektif untuk mendukung deteksi retinopati diabetik secara otomatis dan lebih akurat.

-Diabetic retinopathy is a serious complication of diabetes mellitus that causes damage to retinal blood vessels and can potentially lead to blindness. Early detection through fundus image segmentation is an important step to help diagnose this disease. This study aims to develop a retinal fundus image segmentation model using the R2AU-Net architecture with the application of Contrast Limited Adaptive Histogram Equalization (CLAHE) as pre-processing to increase image contrast. The CLAHE process was carried out with clipLimit 6.0 parameters and tileGridSize 8×8, then the dataset was divided into two categories, namely without CLAHE pre-processing and with CLAHE pre-processing. The model was trained using the Stochastic Gradient Descent optimizer with a learning rate of 10⁻², 100 epochs, and a batch size of 4, and evaluated using the Pixel Accuracy, Dice Coefficient, and Intersection over Union (IoU) metrics. The results show that the model with CLAHE preprocessing produces Pixel Accuracy of 0.8833, Dice Coefficient of 0.4519, and IoU of 0.3440, while without CLAHE it obtains Pixel Accuracy of 0.9261, Dice Coefficient of 0.3742, and IoU of 0.2746. The increase in Dice and IoU values indicates that the application of CLAHE successfully improves the quality of segmentation by highlighting the fine blood vessel structure, so that the combination of CLAHE and R2AU-Net is effective in supporting automatic and more accurate detection of diabetic retinopathy.