<![CDATA[Classification of diabetic retinopathy (DR) based on retinal images is important for early detection, but is often hampered by poor image quality such as noise, uneven lighting, and low contrast. This study analyzes the effect of applying three image filtering techniques, namely Graph Laplacian Filtering (GLF), Graph Convolutional Network (GCN), and Non-Local Means (NLM), on improving the performance of Diabetic Retinopathy classification. The three methods were compared with a baseline model without filtering using VGG16 and evaluated through accuracy, AUC, loss, and image quality metrics such as PSNR, SSIM, MSE, and RMSE.The results showed that graphical and spatial filtering did not always improve classification performance, as VGG16 Fine-Tuning without filtering achieved the highest accuracy of 97.84%. Combinations with NLM, GCN, and Graph Laplacian resulted in lower accuracy due to the smoothing effect that removed important microfeatures on the retina. However, NLM remained effective in reducing noise without disturbing edge structures. These findings confirm that improving image visual quality does not always correlate with CNN accuracy, so preprocessing must focus on preserving diagnostic features. ]]>
Copyrights © 2026
